# ZeroLog
> Your AI. Your Tools. Your Freedom. One workspace. Thousands of models. Zero vendor lock-in. Linux containers with root for Claude Code and Codex, from CHF 9 a month.
Preview: ZeroLog is not open to the public yet. Accounts are created on request.
Pages are server-rendered HTML. The English home page, product pages, pricing, documentation and news articles also have a
Markdown version: add `.md` to the path (the home page is https://zerolog.ch/index.md) or request the page with
`Accept: text/markdown`. German pages have a Markdown version only for the home page (https://zerolog.ch/de.md) and the agents page (https://zerolog.ch/de/agents.md); the other German pages, the status page and the sign-up page have none.
The whole documentation, the home page and the pricing in one file: https://zerolog.ch/llms-full.txt
The signed-in app (/next/) and the API (/api/) are private and not meant for crawling.
## About ZeroLog
ZeroLog is an AI workspace run by Gian-Luca Luongo, a sole proprietor in Wettingen, Switzerland. It gives you a persistent Linux container with root access, a chat with many open-weight models, and Claude Code or Codex on your own subscription. Chats, files and container disks are stored in Seon (Aargau). Plans start at CHF 9 a month; ZeroLog is in a closed preview and accounts are created on request. It is unrelated to the zerolog logging library for Go.
- Agents: Claude Code and Codex run in your container, on your own vendor account or on ZeroLog models. ZeroLog Code is a terminal client based on OpenCode. /goal keeps an agent working until a separate check on the server says the goal is met. By default every command the AI wants to run waits for your approval.
- Data: chats, files and workspace disks are stored in Switzerland. Prompts and attachments go to the model provider, Featherless (USA), which confirmed in writing on 28 September 2026 that it stores no prompt or answer content and does not train on it. That is the provider's commitment; ZeroLog cannot inspect its systems.
- Isolation: each workspace is its own container (gVisor, its own small virtual machine in the VM class), and workspaces cannot reach each other.
- Models: one inference provider, paid from prepaid credit at the provider's rate plus 2 %. A plan buys the machine and tools; credit pays for model usage.
- Not available yet: sharing a container with another person. The Android app and the browser extension are installed by hand.
- Containers start with Mini. Free is an account without a container.
- A /goal run stops after 4 h on Mini, 8 h on Pro and 12 h on Max; Ultra has no limit. The limit is per run, not per month.
- Deep Research starts with Pro. Guest virtual machines (VM class) are Ultra only.
- Compute boost: +2 vCPU · 8 GB · 25 GB disk for CHF 12.00 a month, up to 2 per account, on Pro, Max and Ultra.
- Paying yearly costs ten months instead of twelve: Mini CHF 90.00, Pro CHF 150.00, Max CHF 240.00 and Ultra CHF 790.00.
- Your own AI subscription (for example Claude Code or Codex signed in with your vendor account) runs in a container on any plan that has one; the vendor bills you and ZeroLog credit is not touched. Your own provider API key works on every plan and carries no ZeroLog margin.
Questions and answers in one place: [FAQ](https://zerolog.ch/faq). Prices are in the Pricing section below and on [Pricing](https://zerolog.ch/pricing.md).
---
Private preview
# Your AI. Your Tools. Your Freedom.
The ZeroLog monogram ZL as a glass prism with a floor reflection
One workspace. Thousands of models. Zero vendor lock-in. Claude Code and Codex with root access, from CHF 9 a month.
[Request access](https://zerolog.ch/warteliste) [See pricing](https://zerolog.ch/pricing)
- Data stored in Seon, Aargau
- from CHF 9/month
- your own Linux container
## ZeroLog in six facts
ZeroLog is an AI workspace from Switzerland: a Linux container of your own, an AI agent working in it, many models in one chat, and prices you can read before you sign up.
- [Persistent Linux containers](https://zerolog.ch/#container)
- [Coding agents](https://zerolog.ch/#container)
- [Hosted in Switzerland](https://zerolog.ch/#schweiz)
- [Keys and secrets stay encrypted.](https://zerolog.ch/#tresor)
- [Any model, bring your own key (BYOK)](https://zerolog.ch/#modelle)
- [Transparent prices](https://zerolog.ch/#preise)
## Persistent Linux containers
From Mini
Your own container with root access. Installed packages and files stay across restarts and chats. Pausing keeps the disk. In the container view, one click joins it to your own Tailscale network, and one click opens it over SSH or as a drive on Windows (ZeroLog Connect, not code-signed yet).
### Web IDE
code-server, the VS Code editor, opens in your browser on your container’s files. It sits behind your ZeroLog login, not on a public port.
### Coding agents
Native Claude Code and Codex run in a private workspace with your own provider account. ZeroLog Code uses the ZeroLog catalogue. Each tool keeps its own conversation.
[How containers work](https://zerolog.ch/docs/containers) [Files and editor](https://zerolog.ch/docs/files) [Supported CLIs](https://zerolog.ch/docs/cli)
A glass cube with a glowing terminal prompt inside
## Hosted in Switzerland
Chats, files and workspace disks are stored on our server in Seon (Aargau), Switzerland. Model requests are the exception: your prompts go to the model provider in the USA.
[Where your data goes](https://zerolog.ch/swiss-ai-workspace)
A mountain range made of glass prisms
## Keys and secrets stay encrypted.
What is encrypted, what is not, and where data is stored. The privacy policy and the trust documentation have the full scope.
- **Stored secrets**: Vault entries, provider keys and access tokens are encrypted on the server with AES-256-GCM.
- **Anonymous chats**: An anonymous chat is encrypted in your browser, and the server cannot read it. If you lose the passphrase, the chat is lost. This is off by default.
- **Everything else**: Accounts, chats and container disks are stored in Switzerland. Regular chats are processed on the server, so the service can read them. Model inference runs with providers abroad.
[Trust documentation](https://zerolog.ch/docs/trust) [Privacy policy](https://zerolog.ch/legal)
A key and a padlock made of glass prisms
## Any model, bring your own key (BYOK)
Pick from a catalogue of tens of thousands of open models. Price and context length show per model. Or use your own API key.
[Choosing a model](https://zerolog.ch/docs/models) [Open-source models](https://zerolog.ch/open-source-models) [Your own Claude or ChatGPT plan](https://zerolog.ch/bring-your-own-ai-subscription) [Claude Code in the cloud](https://zerolog.ch/claude-code-cloud)
Five blank glass cards fanned out, tinted from magenta to cyan
## Transparent prices
From **CHF 9** a month
A free account costs CHF 0. Mini, Pro, Max and Ultra each include a container and monthly AI credit. Model usage shows its cost per answer.
Free CHF 0, no container · Mini CHF 9 · Pro CHF 15 · Max CHF 24 · Ultra CHF 79, up to 3 containers. Per month.
- Mini · CHF 9 a month · CHF 2 credit
- Credit is spent on any model at the provider’s rate plus 2 %.
[See all plans](https://zerolog.ch/pricing)
A balance scale made of glass prisms
## The ZeroLog family
Separate tools. A shared identity.
- ### ZeroLog
Safety, Privacy and Quality Combined.
[Details](https://zerolog.ch/)
- ### ZeroLog Code
Your workspace, in your terminal.
[Details](https://zerolog.ch/code)
- ### ZeroLog Desktop
Your workspace, native on Windows.
[Details](https://zerolog.ch/desktop)
- ### ZeroLog Mobile
Your AI. Anywhere.
[Details](https://zerolog.ch/mobile)
- ### ZeroLog Extension
Your browser, with ZeroLog.
[Details](https://zerolog.ch/extension)
## Built to stay. Free to leave.
[Request access](https://zerolog.ch/warteliste) [See pricing](https://zerolog.ch/pricing)
---
Agents that run without you · Switzerland
# AI agents that keep working after you close the tab.
A ZeroLog agent runs in your own Linux container, on the server, not in your browser. You give it a condition that means “done”, and a separate model reads each round and decides whether that condition is met. Every command can wait for your approval, and every run leaves a log with its state and its cost.
[Request access](mailto:info@zerolog.ch?subject=Zugang%20zu%20ZeroLog) [how it knows it is done](https://zerolog.ch/agents#goal)
the stopping rule
## The agent does not decide when it is finished.
Type `/goal` and the condition, for example “pytest reports 0 failed”. After every answer a second, smaller model reads the conversation and rules on one question: is the condition true? If not, it says what is missing, and the agent is called again with that reason. The loop runs on the server, so closing the tab does not stop the next round. In the chat, /goal is limited by rounds, not by the clock.
Write a condition the checker can see in the chat: a test count, a build without errors, a file that has to match. “Make it better” never becomes true.
[/goal in the docs](https://zerolog.ch/docs/goal)
1. agent works in the container
2. checker reads the conversation since the goal was set
3. not met the reason goes back to the agent
stops when
- **met** the condition is true
- **impossible** a blocker only you can clear, such as access you have to grant
- **stalled** the same gap four rounds in a row
- **capped** 25 rounds by default
├ condition up to 4000 characters ┤ not available in anonymous chats
more than one agent
## Split the work, then let two reviewers try to break it.
Switch a chat to `/ultracode` and the model may hand a large task to several agents in your container when that is faster than doing it alone. It decides when; a short question still gets a short answer. After an agent finishes its part, two reviewers go through the files it changed, one looking for what is wrong and one for what is missing. A part they reject goes back for a fix before you see it.
Several agents on one task means several model bills. Use it where a wrong answer is expensive: an audit, a migration, a review before a release.
[/ultracode in the docs](https://zerolog.ch/docs/ultracode)
what it may do without asking
## You decide which commands run unattended.
### Approve each command
Every command the agent wants to run appears in full and waits. Approve runs it once, Deny tells the model you said no, and you can edit the command before it runs. Leave a request for 5 minutes and it expires; the agent stops there and sums up what is left. You can switch a chat to automatic while the agent works; approvals already waiting are released.
[Approvals in the docs](https://zerolog.ch/docs/approval)
### Standing permissions
Grant routine actions once and they run without asking; anything outside a grant waits in a queue for you. Grants are explicit and can be revoked. Moving money is not a permission you can give: it is refused every time.
[Automation in the docs](https://zerolog.ch/docs/automation)
### A report afterwards
Each long run shows up under Jobs with its state and its log. Follow it live or read it the next morning, and see what that particular run cost.
[Jobs in the docs](https://zerolog.ch/docs/jobs)
### On a schedule
Give an agent a task and a time: a nightly build, a daily check, a weekly summary. Schedules follow your timezone, and every run leaves a log.
[Scheduled tasks in the docs](https://zerolog.ch/docs/tasks)
where it runs
## A container you have root on, not a sandbox that forgets.
The agent works in the same Linux container you use yourself: installed packages stay, files stay, and on every paid plan a service it starts keeps answering on a subdomain you pick. Containers run isolated from every other account. A plan sets how big the machine is and how long one agent run may last.
[Containers](https://zerolog.ch/docs/containers) · [Isolation](https://zerolog.ch/docs/isolation) · [Plans](https://zerolog.ch/docs/plans)
per plan
- Free · CHF 0 no container, no agent runs
- Mini · CHF 9 up to 4 hours per run
- Pro · CHF 15 up to 8 hours per run
- Max · CHF 24 up to 12 hours per run
- Ultra · CHF 79 no limit (24/7)
prices a month, prepaid models are paid from credit
what a run costs
## The plan pays for the machine. Each model call is paid from credit.
Model tokens cost the provider’s rate plus 2 %, taken from your ZeroLog credit, and the figure under each reply is the figure you are charged. When the credit reaches zero, paid models stop; nothing is billed afterwards. A /goal run also ends at the time limit of your plan. No plan sets a spend cap per run by default.
[Cost per reply](https://zerolog.ch/docs/cost) · [Credit and top-ups](https://zerolog.ch/docs/credit) · [Limits and what has no cap](https://zerolog.ch/docs/limits) · [/goal and /ultracode](https://zerolog.ch/unattended-agents)
the details
## Everything on this page, one topic at a time.
- [/goal](https://zerolog.ch/docs/goal) conditions, rounds, when a goal stops
- [/ultracode](https://zerolog.ch/docs/ultracode) several agents, two reviewers per part
- [Approvals](https://zerolog.ch/docs/approval) approve, deny, edit, expiry
- [Automation](https://zerolog.ch/docs/automation) standing permissions and the queue
- [Jobs](https://zerolog.ch/docs/jobs) state, log and cost of each run
- [Scheduled tasks](https://zerolog.ch/docs/tasks) agents on a timetable
- [Plans](https://zerolog.ch/docs/plans) machine size and run time
## Hand it the job, and read the log in the morning.
Agent runs need a container, which starts with the Mini plan. Accounts are created on request while ZeroLog is in preview.
[Request access](mailto:info@zerolog.ch?subject=Zugang%20zu%20ZeroLog) [compare plans](https://zerolog.ch/pricing)
---
ZeroLog
# Frequently asked questions about ZeroLog
Plain answers on what ZeroLog is, where your data lives, which agents and models run, what the plans cost and what the preview cannot do yet. Each answer points to the documentation behind it.
[See pricing](https://zerolog.ch/pricing) [Read the docs](https://zerolog.ch/docs)
## The product
- ### What is ZeroLog?
ZeroLog is a platform for AI agents with persistent Linux workspaces in the cloud. You chat with a model, attach a container (a real Linux machine with root access) and let agents build and run your software there. It is hosted in Seon (Aargau), Switzerland, and currently in a closed preview.
- ### Is ZeroLog related to the zerolog logging library for Go?
No. The name is a coincidence. ZeroLog at zerolog.ch is a hosted platform for AI agents. The zerolog library for Go is an unrelated logging library.
- ### How do chat, container, plan and credit fit together?
A chat is the conversation and uses one model at a time. A container is a real Linux machine that exists independently of any chat, and its files persist. The plan sets the hardware and tools you may use, and paid plans add a monthly credit allowance. Credit is prepaid money for model usage; it never buys hardware. See [How it fits together](https://zerolog.ch/docs/concepts).
- ### Which agents and tools run in a container?
Claude Code and Codex run in a private workspace with your own eligible provider account. [ZeroLog Code](https://zerolog.ch/code) uses ZeroLog models. In ZeroLog chat, [/goal](https://zerolog.ch/docs/goal) sets a finish condition and /ultracode allows delegation. Native CLI commands and permissions are separate. AI actions wait for approval by default; review the mode before a long run.
- ### Which models can I use?
Models for customers come from one inference provider, Featherless in the USA. The picker lists each model with its price per million tokens and its context length, and you can switch models in the middle of a chat. A model can be unavailable at times because of provider capacity. See [Choosing a model](https://zerolog.ch/docs/models).
- ### Is there an app, a terminal client or a browser extension?
Yes, all three. [ZeroLog Code](https://zerolog.ch/code) is the terminal client. [ZeroLog Mobile](https://zerolog.ch/mobile) is a native Android app. The [browser extension](https://zerolog.ch/extension) adds a side panel to Chrome and Firefox. The app and the extension are installed by hand from a file for now.
## Data and privacy
- ### Where is my data stored?
Chats, files and workspace disks are stored on servers in Switzerland. To generate a reply, your prompts and attached files are sent to the model provider, Featherless in the USA. Further recipients, such as Stripe for payments and Cloudflare for reaching the site, are listed on the [trust page](https://zerolog.ch/docs/trust) and in the [privacy policy](https://zerolog.ch/privacy).
- ### Does the model provider keep my prompts or train on them?
Featherless confirmed in writing on 28 September 2026 that it does not store the content of prompts or answers, keeps only metadata for abuse monitoring and error handling, and does not train on your content. That is the provider’s commitment; ZeroLog cannot inspect its systems. Each answer shows whether [zero data retention](https://zerolog.ch/docs/zdr) was enforced, stated by the provider, or not guaranteed.
- ### How are workspaces isolated?
A workspace starts as its owner’s container. Other users need an explicit sharing grant with read-only or read-write access. Personal native CLI credentials need a private workspace. Containers run on gVisor and outbound traffic is filtered. See [isolation](https://zerolog.ch/docs/isolation), [sharing](https://zerolog.ch/docs/share) and [outbound network](https://zerolog.ch/docs/egress).
- ### Can I export or delete my data?
You can download any file or a whole folder of a container as an archive, and each workspace exports as a Dockerfile and a Compose file. A paused container keeps its disk until you delete it. Deleting your account removes the data linked to it. Backups follow their own retention, and some older audit-log entries keep an account reference. The [privacy policy](https://zerolog.ch/privacy) describes this in detail.
- ### How do I sign in?
With a username and password. An email address is optional, but without one you cannot reset your password by email. Sign-in runs through ZeroLog’s own Keycloak service on servers in Switzerland, and you can add an authenticator app, a passkey or email codes as further factors. See [Two-factor authentication and passkeys](https://zerolog.ch/docs/mfa).
## Plans and credit
- ### What plans are there and what do they cost?
There are five plans, billed in Swiss francs: Free (CHF 0), Mini (CHF 9 a month), Pro (CHF 15), Max (CHF 24) and Ultra (CHF 79). Paying yearly costs ten months instead of twelve. Mini to Ultra include a Linux container. Model credit each month: Mini CHF 2, Pro CHF 4, Max CHF 8, Ultra CHF 30. Current prices and limits are always on the [pricing page](https://zerolog.ch/pricing).
- ### What does the Free plan include?
Yes, as an account without a container. Free gives you chat, 1 GB of file storage and the secrets vault. You pay for models with your own AI access or with credit you buy. Containers and the tools that run in them start with Mini.
- ### Which plan do I need for agent runs, Deep Research or virtual machines?
A /goal run can last up to 4 hours on Mini, 8 hours on Pro and 12 hours on Max; on Ultra it has no limit. The limit applies to each run, not to the month. Deep Research starts with Pro. Guest virtual machines (VM class) are Ultra only, and Ultra allows up to three containers that share one total. See [Plans](https://zerolog.ch/docs/plans).
- ### What happens when my credit runs out?
Paid models stop, and nothing is charged afterwards. On any plan you can subscribe to a monthly credit pack, billed by card through Stripe and renewing until you cancel: S adds CHF 5.50 for CHF 5.90, M adds CHF 15.00 for CHF 15.90 and L adds CHF 50.00 for CHF 51.50. Models cost the provider’s rate plus 2 %, and the price you see already includes it. Included monthly credit is spent first and expires three months after it was added. See [Credit and top-ups](https://zerolog.ch/docs/credit).
- ### Can I use my own AI subscription or API key?
Yes. With a container (Mini and up) you can sign in a command-line tool such as Claude Code or Codex with your own vendor account, for example a Claude or ChatGPT subscription. The vendor bills you, nothing reaches your ZeroLog credit, and the vendor’s own limits apply. You can also add your own provider API key; that usage goes to your account there and ZeroLog adds no margin. See [Your own vendor subscription](https://zerolog.ch/docs/byoagent) and [Your own API key](https://zerolog.ch/docs/byok).
- ### What is the compute boost?
On Pro, Max and Ultra you can add a boost: 2 vCPU, 8 GB RAM and 25 GB SSD for CHF 12 a month, up to two boosts per account. Mini has no boost.
## Access and preview
- ### Can I sign up today?
ZeroLog is in a closed preview, and new accounts need an invitation. Request access on the [waitlist](https://zerolog.ch/warteliste) and confirm your email address through the link you receive. Plans and prices are published, but accounts are created on request.
- ### What can the preview not do yet?
Container owners on plans with sharing can invite users with read-only or read-write access; see [sharing and credential restrictions](https://zerolog.ch/docs/share). Android is a pre-release build. Outbound mail and some service ports are closed in workspaces. An unavailable customer model shows an error instead of silently switching models.
- ### Who runs ZeroLog and how do I reach support?
ZeroLog is run by a sole proprietor in Switzerland; the imprint has the details. Write to info@zerolog.ch for support, access requests and security reports. Vulnerability reports follow [security.txt](https://zerolog.ch/.well-known/security.txt) (RFC 9116), and changes to the service are announced in the [Newsroom](https://zerolog.ch/news).
## Apps and clients
- ### What is ZeroLog Desktop?
ZeroLog Desktop is the ZeroLog app for Windows, currently version 0.4.2. It puts your chats, a coding agent for a local folder or a cloud container, jobs, and your containers and Windows VMs in one window. With one click you can mount a container as a Windows drive or open SSH into it. It uses the same account and data as the web and updates itself. It needs Windows 10 (1809 or later) or 11, 64-bit. See [ZeroLog Desktop](https://zerolog.ch/desktop).
- ### Is ZeroLog Desktop built on Electron?
Yes. ZeroLog Desktop is built with Electron since version 0.3.0, which replaced the earlier Qt preview (0.2.1). Version 0.4.0 brought in the functions of the earlier native Windows app, and 0.4.1 added passkeys and account deletion. Installed 0.3.0, 0.4.0 and 0.4.1 apps update themselves. See [ZeroLog Desktop in the docs](https://zerolog.ch/docs/zerolog-desktop).
- ### Does ZeroLog Desktop run on macOS or Linux?
No. There is no build for macOS or Linux. ZeroLog Desktop is for Windows 10 (1809 or later) and Windows 11, 64-bit, and the installer refuses older systems. On macOS or Linux you can use the web app, the browser extension or the terminal client [ZeroLog Code](https://zerolog.ch/code), which installs on Linux, macOS and Windows.
- ### What is ZeroLog Code and how do I install it?
ZeroLog Code is a terminal client based on OpenCode. Install it with one command, curl on Linux and macOS or PowerShell on Windows, then run zerolog login and confirm in your browser. After that your local files stay in step with your workspace. The steps are on the [ZeroLog Code page](https://zerolog.ch/code).
- ### Does ZeroLog Code use my ZeroLog credit?
Yes, when the run uses a ZeroLog model. Those runs count against your account limits and appear with your other usage, and at zero credit paid models stop. Command-line tools you sign in with your own vendor account are billed by the vendor instead, and the vendor’s limits apply. See [Claude Code and other CLIs](https://zerolog.ch/docs/cli).
- ### Can I follow a ZeroLog Code run in the web chat?
Yes. Console activity and command output from ZeroLog Code appear in the web chat while they happen, so you can watch a run from the browser as well as from the terminal. Commands the AI wants to run still wait for your approval by default.
- ### Is ZeroLog Code the same as zerocode?
Yes. The client was called zerocode before. The installer replaces an older copy where it finds one, and the old command zerocode keeps working, so existing scripts continue to run. Each release lists its SHA-256 sums in checksums.txt, so you can verify a download.
- ### Can an AI agent run on my Android phone?
Yes. The native [ZeroLog Mobile](https://zerolog.ch/mobile) app (Android 8.0 or later, pre-release build) runs an agent on the phone itself, with Python 3.11, a shell and app control. Each action asks first unless you turn approvals off. A folder you pick can stay in step with your container. See [Android app](https://zerolog.ch/docs/app).
- ### How do I keep an AI agent from using up my credit?
Your credit is the hard stop: at zero, paid models stop and nothing is billed afterwards. A /goal run also ends at your plan’s time limit, and delegated sub-agents are capped at CHF 0.50 per chat turn. There is no spend cap per run, chat or agent, so size your credit to the job. See [Spend limits](https://zerolog.ch/docs/limits).
---
# Pricing
> Five plans. Free is an account without a container: bring your own AI accounts or buy credit. Mini (CHF 9.00 a month), Pro (CHF 15.00), Max (CHF 24.00) and Ultra (CHF 79.00) add the machine and a monthly model credit. Credit buys models at the provider's rate plus 2 %. Prepaid in CHF, hosted in Switzerland. Every plan stops at its limit, and nothing is billed after the fact.
Preview: ZeroLog is not open to the public yet. Accounts are created on request (info@zerolog.ch).
## Plans
Free has no container, no autonomous runs and no Deep Research. Mini, Pro and Max have one container each; Ultra has up to three that share one total. A /goal run stops after 4 h on Mini, 8 h on Pro and 12 h on Max (Ultra: no limit). Deep Research starts with Pro. Yearly billing is ten months for the price of twelve (Mini CHF 90.00, Pro CHF 150.00, Max CHF 240.00, Ultra CHF 790.00).
| Plan | Per month | Machine | Credit each month | Containers | Ports | Agent run |
| --- | --- | --- | --- | --- | --- | --- |
| Free | CHF 0 | no container | - | 0 | none | none |
| Mini | CHF 9.00 | 1 vCPU · 4 GB · 10 GB disk (all containers together) | CHF 2.00 | 1 | 1 | 4 h |
| Pro | CHF 15.00 | 2 vCPU · 8 GB · 15 GB disk (all containers together) | CHF 4.00 | 1 | 1 + 1 domain | 8 h |
| Max | CHF 24.00 | 4 vCPU · 12 GB · 25 GB disk (all containers together) | CHF 8.00 | 1 | 2 + 1 domain | 12 h |
| Ultra | CHF 79.00 | 12 vCPU · 48 GB · 150 GB disk (all containers together) | CHF 30.00 | 3 | 8 + domains | unlimited |
## Credit
Credit packs add model credit on any plan and are spent on any model in the catalogue at the provider's rate plus 2 %. What you do not use carries forward, up to 3 months' worth.
| Plan | Price | Credit | Carries up to | Container |
| --- | --- | --- | --- | --- |
| Credit S | CHF 5.90 | CHF 5.50 | CHF 16.50 | none (any plan) |
| Credit M | CHF 15.90 | CHF 15.00 | CHF 45.00 | none (any plan) |
| Credit L | CHF 51.50 | CHF 50.00 | CHF 150.00 | none (any plan) |
A compute boost adds 2 vCPU, 8 GB RAM and 25 GB disk for CHF 12.00 a month (up to 2 per account, on Pro, Max, Ultra).
Topping up by hand instead carries no fee, only a 3 % currency buffer, quoted before you pay. Bring your own provider key and the margin is zero.
## How credit behaves
- The price under a reply already includes the 2 %, so the figure you see is the figure you are charged.
- Subscription credit and topped-up credit sit in one balance; subscription credit is spent first.
- Unused subscription credit carries forward up to 3 months' worth.
- At zero the paid models stop. Nothing is charged after the fact.
- Change plan any time; the part-month is settled on the next invoice.
- Cancel and topped-up credit stays spendable. Carried subscription credit expires 90 days after the last paid month.
- Traffic on your own provider key goes to the provider at their rate; ZeroLog takes nothing on it.
---
# First 5 minutes
Five steps, in order. Two build a chat, one builds a machine, and the fourth attaches the machine to the chat. The fifth is where the AI stops and asks you before it runs anything.
**The interface starts in German.** Every label has an English version: switch with the language button in the top bar, or under Settings → General → Language. This documentation quotes the English labels throughout, so switch first if you want them to match what you see.
1. ## 01Request access
New accounts currently need an invitation. Write to [info@zerolog.ch](mailto:info@zerolog.ch), use [Request access](https://zerolog.ch/pricing) (it opens an email), or join the [waitlist](https://zerolog.ch/warteliste). An administrator creates the account, so there is no public sign-up form yet. You then sign in on the sign-in page. An email address is optional; add or confirm it under Settings → Sign-in & security.
A confirmed address can receive password reset links ("Forgot password" on the sign-in page) and security alerts. Without one, you cannot reset your password by email. A passkey can also sign you in without the password.
2. ## 02Start a chat
Choose a model shown in the picker, then ask it something. The picker shows the model's price and context length.
3. ## 03Create a container
Open Compute and create a container. The default image is Linux with Ubuntu 24.04. Free accounts have no container; Mini and the larger plans do.
4. ## 04Attach it to the chat
In the chat's Workspace panel, choose Link a workspace. Now the AI works inside that machine.
5. ## 05Ask it to build something
The first command it wants to run stops and waits for your approval, unless it is on your list under Settings → AI → Allowed commands.
**Check the price before you send.** You can change models from the picker at any time.
---
# How it fits together
Four things, and it helps to keep them apart.
## The four things
You use
Models
### Chat
The conversation. It belongs to you.
one model at a time, changeable
Hardware
### Container
A real Linux machine, independent of any chat. You can attach the same container to several chats.
its files persist
You pay
Models
### Credit
Prepaid money for model usage. Separate from the plan, though every paid plan adds some each month (see [Plans](https://zerolog.ch/docs/plans) and [Credit](https://zerolog.ch/docs/credit)).
paid per request
Hardware
### Plan
What hardware you may use.
- cores
- memory and disk
- how many containers run at once
- how long an agent may run
- how many public ports
## The important consequence
Plan Model usage
A plan does not pay for model usage. It only adds some credit each month.
Credit Hardware
Credit does not buy hardware.
They are two meters because they are two different costs.
## What survives closing the tab
*Keeps running*
- A chat, an agent or a [/goal](https://zerolog.ch/docs/goal) run once you have sent it
- A [terminal](https://zerolog.ch/docs/terminal) session you can reattach to later
- A [container](https://zerolog.ch/docs/containers)'s files, across restarts
*Needs your action*
- Stopping a run: the stop button or an explicit abort
- Nothing stops it just because the tab closed or the page reloaded
See [Isolation](https://zerolog.ch/docs/isolation) for what else does and does not change about a workspace.
## Chat, container, plan and credit
Mermaid source of the diagram on this page:
```mermaid
%% How chat, container, plan and credit relate. Facts from server/tiers.js (hardware per plan), server/plans.js (monthly credit grant, 3 month expiry), web kontext-leiste.tsx (Link a workspace).
flowchart LR
plan["Plan
sets the hardware"]
chat["Chat
one model at a time"]
credit["Credit
prepaid, for models"]
box["Container
a Linux machine,
files persist"]
req["Model request
paid per request"]
plan -->|"cores, memory, disk,
run length, ports"| box
chat -->|"attach; one container
can serve several chats"| box
chat -->|"each answer"| req
credit -->|"pays"| req
plan -->|"adds some credit
each month"| credit
linkStyle 3 stroke:#8f7ac9
```
---
# Waitlist email links
Opening a waitlist email link only shows a confirmation page. Your entry changes only after you choose an action and submit the form.
## Confirm your email
Open the confirmation link and choose **Anmeldung bestätigen** (Confirm registration). The page tells you when confirmation is complete, or that the address was already confirmed. The link works for 24 hours. If you never click it, nothing happens and the address is removed at the next cleanup after seven days from your submission. If the link has expired, sign up again on the [waitlist page](https://zerolog.ch/warteliste) to get a new email.
## Leave the waitlist
Open the unsubscribe link and choose **Meinen Eintrag entfernen** (Remove my entry). The page confirms the request without revealing whether the address was on the list.
The confirmation email carries this unsubscribe link and also a one-click unsubscribe header. If your email app shows that option, using it also removes your entry.
## What a waitlist email link does
Mermaid source of the diagram on this page:
```mermaid
%% Waitlist email links. Facts from server/routes/warteliste.js (GET and POST /warteliste/bestaetigen, /warteliste/abmelden) and server/warteliste.js (24 h link, 7 day deletion).
flowchart TB
m["Email from the waitlist
with a confirm link and
an unsubscribe link"]
g["You open a link
(GET)"]
p["A page asks:
Bestätigen or Abmelden
Nothing has changed yet"]
b["You press the button
(POST)"]:::hot
ok["Entry confirmed
or removed"]
bad["Link expired or invalid:
sign up again on
the waitlist page"]
scan["Mail scanners and
link previews only GET"]
one["One-click unsubscribe
from your email app
(POST)"]
m --> g --> p --> b --> ok
scan --> p
b -->|"after 24 hours"| bad
m --> one --> ok
classDef hot stroke:#8f7ac9
```
---
# Chat basics
Type, send, read.
The reply streams token by token, and the cost strip under it fills in while it writes.
## Edit and send again
Your edit goes out as a new message at the end of the thread.
History is not rewritten.
## Fork
Copies the thread into a new chat: up to a message you pick, or the whole chat from the chat menu.
The original stays.
## Stop
Halts generation. The partial reply stays in the chat.
You are billed for what was produced.
## Rename
Give a chat a name of your own.
Or let the first 60 characters of your first message title it.
## Export
Take a chat out as Markdown or JSON.
From the chat menu.
## Pause
Stops the running reply and pauses the containers attached to the chat.
Nothing is deleted. Resume checks your plan again.
## What each action does to a chat
Mermaid source of the diagram on this page:
```mermaid
%% Chat basics: what each action does to a thread. Facts from web/apps/web/components/chat/message-list.tsx, chat-menu.tsx, chat-actions.tsx, server/forks.js, server/routes/chats.js (paused, export, fork).
flowchart LR
reply["Reply streams in,
cost strip fills
while it writes"]
stopped["Partial reply
stays in the chat"]
edit["Your edit goes out as a
new message at the end"]
child["New chat: copy of the
thread up to that message,
titled with (Fork)"]:::hot
paused["Running reply stops,
attached containers pause,
nothing is deleted"]
file["Download:
Markdown or JSON"]
reply -->|Stop| stopped
reply -->|Edit and send again| edit
reply -->|Fork from a message| child
reply -->|Pause| paused
reply -->|Export| file
linkStyle 2 stroke:#8f7ac9
```
---
# Choosing a model
The picker above the input lists the models available to your account. Every row carries the same two numbers.
Every row shows
- **Price**: per million tokens
- **Context**: length
Switch mid-conversation and the thread comes with you.
Kinds
## Three kinds worth telling apart
Costs little
### Low-cost models
They cost little per request. The picker shows each model's price and context length, so you can compare before you choose.
Costs more
### Reasoning models
They think before answering. Worth it on hard problems, not on simple ones.
Longer context
### Long-context models
They matter when you paste a big file or work a long thread.
Where the catalogue has them, a row also shows a class label (a rough family tag) and a concurrency cost: how many of the shared inference slots one request to that model uses. Heavier models use more, which is why they can hit a busy limit sooner under load.
Full catalogue
## Recommended picks, or search everything
The picker has two views: **Recommended**, a short curated list, and **All Featherless models**, which searches the complete catalogue (tens of thousands of models). In the full view you can sort and filter by model type, and **Load more** brings further results.
ZeroLog keeps the previous model list when an incomplete Featherless catalogue would hide many models. Small retirements and returning models still update normally.
For more control, open **Models** in the navigation. The same toggle is there, plus filters for class and minimum context, sorting (context and price, popularity, downloads, newest, model size, price, context, name), a model-type filter, and tags such as vision, reasoning and long context so you can spot what a model is built for at a glance.
In practice
## Cheap to explore, strong for the hard step
One pattern that holds up:
1. 01
Cheap model to explore.
2. Switch 02
Strong model for the hard step.
3. Switch back 03
Cheap again to tidy up.
Defaults
## Default model and automated tiers
New chats start on your default model from Settings; without one, ZeroLog uses its system default. Automated background work, such as the checks and summaries inside Jobs and research runs, uses three tiers: **everyday**, **medium** and **complex**. They follow ZeroLog's current configuration, not your picker.
If a model fails
## When the model you picked cannot answer
If the selected model cannot handle a request, the chat shows an error. Check the message, then retry or choose another model shown in your picker.
Malformed request
You see the error; another model may reject it too.
Device tools
The CLI engines (Claude Code and Codex) cannot call the tools of a connected device such as the Android app. The thread says so when it happens.
Availability
## Use models shown in your picker
ZeroLog checks model access across chats and automated features. Agents, Ultracode, automatic model choice and usage limits use models available to your account.
When you explicitly request a model that is not available to you, ZeroLog returns `model_not_available`. Choose a model from your picker and try again.
Featherless models are available to existing accounts. See [Model availability](https://zerolog.ch/docs/provider-pause) for the current access policy and the earlier pause.
Providers sometimes retire models, so the list and prices in your picker can change. In ZeroLog Code, a saved model choice that has been retired may be routed to a replacement. The response identifies both models and includes a notice, whether it streams or arrives all at once. Choose a current model in the picker to set a different one.
Price
## What the quoted price already includes
On ZeroLog credit
Every price you see already includes the service margin (2% at the shipped rate), so what is quoted is what is charged.
On your own key
Usage on [your own API key](https://zerolog.ch/docs/byok) carries no margin at all, because that money never passes through here.
Plan prices and credit packs: [Pricing](https://zerolog.ch/pricing). Model changes are announced in the [Newsroom](https://zerolog.ch/news).
One model in detail: [MiniMax M3 in your coding agent](https://zerolog.ch/minimax-m3-coding-agent).
## How a model is chosen
Mermaid source of the diagram on this page:
```mermaid
%% Choosing a model: picker views and what happens when a model is not available to you. Facts from web/apps/web/components/model-picker.tsx, app/(app)/catalog/page.tsx, server/kundenmodelle.js (model_not_available), server/pricing.js (margin).
flowchart TB
pick["Model picker above the input"]
rec["Recommended:
short curated list"]
all["All Featherless models:
search, sort, model type,
Load more"]
avail{"Available to
your account?"}
use["Chat uses the model.
Price shown includes
the service margin"]:::hot
err["model_not_available:
pick a model from your picker"]
pick --> rec
pick --> all
rec --> avail
all --> avail
avail -->|yes| use
avail -->|no| err
linkStyle 4 stroke:#8f7ac9
```
---
# Model availability
The temporary customer inference pause ended on 14 October 2026. Featherless models are available to existing ZeroLog accounts.
## Model access
Choose an available model in the chat model picker. Individual models can still be unavailable because of provider capacity or account limits. The picker and the error shown for a request describe the current availability. If the pause is switched on again, the picker shows a notice and Featherless models stay unavailable to customer accounts until it ends.
## Your account
New account creation remains restricted to invitations. Enabling model access does not change your existing credit balance and conversation history.
## Paid time
During the earlier pause, paid subscription time and credit expiry were extended by the pause duration. Check Billing for your current plan dates and balance.
## More
See [Choosing a model](https://zerolog.ch/docs/models) for model access and [Privacy](https://zerolog.ch/legal#privacy) for provider processing and retention information.
---
# Files and images in chat
Attach with the clip, or paste an image straight into the box. Up to 8 files go with one message.
Files uploaded to a chat may be checked for known malware. This check looks at the file, not your chat messages. A clean result cannot guarantee that a file is safe.
- **Text files**: are read into the message
- **Images**: go to vision-capable models
- **Other files**: go to the container
## It stays with the chat
Attachments are stored with the chat, so the model can refer back to them later.
## Large files stay in the container
A text file over about 100,000 characters is not put into the message. It is kept [in the container](https://zerolog.ch/docs/containers), and the model gets its path, line count and an outline, then reads the parts it needs. That is cheaper, because it does not read the whole file every turn. Files that are not text, such as PDFs or Office documents, go to the container too, and the model gets only their path. You can also put any file there yourself.
## Sent with the message
Pasting a long text and pressing Enter at once sends the file with the message. Sending reads the live list of attachments, not the last drawn page.
## In research and Ultracode runs
Deep Research and [Ultracode](https://zerolog.ch/docs/ultracode) runs take the files of the chat along as an attached brief. A file too large for the prompt appears as an outline, its first lines and its path under /work/uploads, and the agents read the rest in the container. Images are named but not described, because the planning models read text only. Only files of your own chat are used.
*What the model reads each turn*
### Attached to the chat
Turn 1
Turn 2
Turn 3
The whole file, every turn.
### Read from the container
Turn 1
Turn 2
Turn 3
Only the part it needs.
The bars sketch the idea. They are not measured sizes.
## Where an attachment goes
Mermaid source of the diagram on this page:
```mermaid
%% Files and images in chat: where an attachment goes. Facts from web/apps/web/components/chat/attachments.ts, server/attachments.js (INLINE_MAX, MAX_ATTACH_COUNT, grossdateiBlock).
flowchart TB
file["You attach a file
or paste an image"]
img2["Scaled to 1536 px,
sent with the
message"]
small["Text goes in
the message"]
big["Kept in the
container"]:::hot
path["Stored in the
container, only the
path is sent"]
parts["Model gets path,
lines and outline,
reads parts"]:::hot
file -->|Image| img2
file -->|Text, up to
100,000 chars| small
file -->|Text, over
100,000 chars| big
file -->|Other file| path
big --> parts
linkStyle 2 stroke:#8f7ac9
linkStyle 4 stroke:#8f7ac9
```
---
# Long conversations
A long thread does not have to end. When it comes up against the model's context limit, the older turns are compacted into a summary and the conversation carries on instead of failing.
Fold
Older turns
Compacted
**Summary**
*Hollow rows are the turns the model now reads as one summary. Solid rows it still reads in full.*
Marker
## The seam stays visible
Compaction is not silent. It is shown in the thread, so you can see where it happened and which part of the history has been folded up. The original messages stay stored, and the marker can be expanded to show them.
Cost
## Summarised for everyday use
A model set aside for everyday tasks makes the summary, and the call is billed to you like any other. The latest turns stay in full: six of your turns for automatic compaction, the newest one for Compact now.
When
## Automatic, with a switch
Compaction is on by default for each chat. It starts in the background once the thread passes 70% of the model's window or about 150,000 tokens, whichever comes first. Past 90% it runs before your next message is sent. In the context meter you can switch **Compact automatically** off or press **Compact now**, which needs at least two turns. Paused and anonymous chats are not compacted.
Container
## Nothing in your container moves
Compaction touches the conversation and only the conversation. Files are not part of the context in the first place, so everything in the container is left exactly as it was.
Context length
## Measured against your selected model
The context limit depends on the model selected for the chat. Check its context length in the picker. A thread that exceeds it may need compaction before you continue.
## How a long thread is compacted
Mermaid source of the diagram on this page:
```mermaid
%% Long conversations: when and how a thread is compacted. Facts from server/compact.js (threshold 0.70, hard 0.90, trigger 150000 tokens, keep 6 turns, manual keep 1, sweeper, paused chats) and web/apps/web/components/chat/context-meter.tsx.
flowchart TB
grow["Thread grows toward
the model's context limit"]
soft["Over 70% of the window
or about 150,000 tokens:
background compaction"]:::hot
hard["Over 90%: compaction runs
before your next message"]
now["Compact now
(context meter)"]
sum["Everyday model writes
a summary of the older turns"]
keep["Latest turns stay in full,
summary replaces the rest
in what the model reads"]
mark["Marker in the thread,
originals stay stored"]
grow --> soft --> sum
grow --> hard --> sum
now --> sum
sum --> keep --> mark
linkStyle 0 stroke:#8f7ac9
linkStyle 1 stroke:#8f7ac9
```
---
# Prompt builder
Build a clearer instruction, then decide whether to copy it or use it in a chat. In the app this page is called **Prompt builder**.
## The pass
1. Rough
Write a rough instruction and press **Sharpen**.
2. Sharpened
Press **Ask me a question** to be asked clarifying questions, one at a time, with no fixed limit. Answer or skip, then review and edit the improved version. It lists what changed and the assumptions it made.
3. Use it
Copy it, or send it to a chat.
## Start with a clear instruction
The result is not sent until you choose to send it, unless you switch on sending without asking. It can help with instructions you want to get right before using:
- A review checklist
- A commit-message format
- A [scheduled task](https://zerolog.ch/docs/tasks)'s instruction
## What it learns and what it costs
The builder keeps short notes about your preferences, up to 40. You can read, edit and delete each one, or delete them all. Every pass runs on the everyday model against its own budget per session; when that budget is spent it stops and the text stays usable. Two switches control it: sharpen every message before it goes out, and send the sharpened version without asking.
## From rough text to a sharpened prompt
Mermaid source of the diagram on this page:
```mermaid
%% Prompt optimiser (Prompt builder): from rough text to a sharpened prompt. Facts from web/apps/web/app/(app)/prompt/page.tsx and server/optimizer.js, server/routes/optimizer.js.
flowchart TB
rough["Rough instruction"]
sharp["Sharpen:
suggestion with
changes and assumptions"]:::hot
ask["Ask me a question:
answer or skip,
no fixed limit"]
edit["You review and edit"]
use["Copy it, or send it to a chat"]
rough --> sharp
rough --> ask
ask -->|fold answer in| sharp
sharp --> edit --> use
linkStyle 0 stroke:#8f7ac9
```
---
Reusable instructions
# Skills
Some workflows you end up explaining the same way every time. Write it down once as a skill, switch it on, and it is part of the chat from then on.
## The skill
A skill is a named set of instructions. ZeroLog ships a set of built-in skills, and you can add your own under **Settings, Skills** with a name, a description and the instructions. Your skills are visible to you only; built-in ones are read-only. To change one of yours, add it again under the same name.
## How it reaches the model
A **short skill** is placed in the instructions of every turn. A **long skill** is listed by name and description only, and the model reads the full text when your task touches it. Anonymous chats do not use skills.
## Where it applies
In the app you switch a skill on as a default under Settings, Skills.
### New chats
The skills you switched on are attached to every chat you create afterwards.
### Existing chats
They keep the skills they had when you created them.
## How a skill reaches the model
Mermaid source of the diagram on this page:
```mermaid
%% Skills: from a switch in Settings to the model's instructions. Facts from server/skills.js (skillPlan, SKILL_EINBETTEN_MAX, skill.lesen), server/routes/chats.js (default_skills at chat creation), web/apps/web/components/settings/skills-panel.tsx.
flowchart TB
set["Settings, Skills:
switch a skill on
or add your own"]
chat["New chat is created
with these skills attached"]
turn["Each turn"]
short["Short skill:
full text goes into
the instructions"]:::hot
long["Long skill: only name and
description are listed"]
read["Model reads the full text
when your task needs it"]
set --> chat --> turn
turn --> short
turn --> long --> read
linkStyle 2 stroke:#8f7ac9
```
---
# Memory and writing style
Facts worth carrying between conversations, plus an optional writing-style profile so replies sound the way you want.
## What carries over
**one conversation** **the next one**
- How you like code formatted
- What stack you use
- What you are working on
- Your writing style optional
carried across
## Off until you switch it on
Memory is **off** by default. You switch it on per chat, and it is never available in anonymous chats. When it is on, ZeroLog proposes facts to remember; an entry is kept only after you **confirm** it, and it records the chat it came from. The AI can also search your confirmed entries, propose one itself or delete one when you ask; its proposals wait in Settings, Memory, marked as a suggestion, until you confirm them. The switch is in the chat's settings, or type `/memory on` or `/memory off`.
## Encrypted per account
Entries are stored **encrypted per account** on the ZeroLog server in Switzerland (AES-256-GCM, one key per account). This is **not end-to-end**: the server holds the key, and the model receives the text in plain form when it is added to a chat. When memory is on for a chat, your entries go into every turn and the chat says so. The agent memory your coding agents keep is stored the same way.
The optional **encrypted memory** (Settings, Memory, Encrypted memory) is separate. Your browser encrypts each entry before upload, so the server stores only ciphertext and the key stays on your device. This protects storage only and is **not end-to-end**: while it is unlocked, matching entries (up to about 1,500 characters) are decrypted in your browser and sent in plain text with the message, through the server to the model provider, but only when memory is on for that chat. The chat shows how many entries went along, and a switch in the same settings panel (kept per browser, on by default) stops the sending without locking. “Suggest memories from a recent chat” sends the last 30 messages to a model. The server sees how many entries you have, when they changed and how large they are. Lose both passphrase and recovery code and the entries are gone. Setup never overwrites an existing key (replacing it is an explicit reset that deletes all entries), the passphrase needs at least 12 characters, and the key is cleared when you sign out.
## Never between accounts
your account
any other account
*Memory is yours alone. It is never shared between accounts.*
## Read, edit, delete
You can **read**, **edit** and **delete** every entry under Settings, Memory, Memory & Writing Style. An entry is a fact or a standing instruction; you can also add one by hand or with `/remember`. Nothing is stored that you cannot see. Limits: 200 entries and 20 writing samples (each at least 40 characters) in the normal memory, 2,000 entries in the encrypted one.
## What memory stores and what reaches the model
Mermaid source of the diagram on this page:
```mermaid
%% Memory: two stores and what reaches the model. Facts from server/memory.js (per-chat switch, proposals, MAX_MEMORIES), server/zkmemory.js, web/apps/web/lib/zkmem.ts, server/routes/chats.js.
flowchart TB
chat["Memory switched on
for this chat"]
subgraph ST["Two stores, both yours alone"]
direction LR
plain["Memory: encrypted
per account on the server,
server holds the key"]
zk["Encrypted memory (optional):
your browser encrypts,
key stays on your device"]
end
turn["Matching entries are added
to the turn in plain text,
the chat shows a notice"]:::hot
model["Model provider
receives the text"]
prop["ZeroLog proposes new facts,
kept only after you
confirm each one"]
chat --> ST
ST --> turn --> model
chat --> prop
linkStyle 1 stroke:#8f7ac9
linkStyle 2 stroke:#8f7ac9
```
---
# Context stitching
Bring relevant parts of your own material into a conversation. Switch on sources in Settings → Memory → Context Index.
## The seam
### Your side
- Your past chats
- Gmail, Drive and Calendar, once your Google account is connected
### The conversation
On every message, ZeroLog searches the index by keyword and adds the best four matches. The chat shows a notice naming the sources it used.
*Only your own material is indexed, and no administrator can read it.*
## Where the seam ends
Nothing is indexed until you switch a source on, and each source is separate. Then press Index now.
Search the index, delete one source or delete everything in Settings → Memory → Context Index. Anonymous chats are never indexed. Search is by keyword, not by meaning.
You control what is included
## How the context index reaches a chat
Mermaid source of the diagram on this page:
```mermaid
%% Context stitching: from sources to a chat turn. Facts from server/stitch.js (SOURCES, CHUNK_CHARS 1200, contextFor k=4, lexical FTS) and server/routes/chats.js (stitched_context notice).
flowchart TB
src["Sources you switch on:
your chats, Gmail, Drive, Calendar
(Google needs a connection)"]
idx["Index now: text is cut into
chunks and stored encrypted;
search is by keyword"]
msg["You send a message"]
find["Best 4 chunks for
your message are found"]:::hot
add["Added to the turn,
chat shows a notice
naming the sources"]
src --> idx
idx --> find
msg --> find
find --> add
linkStyle 1 stroke:#8f7ac9
linkStyle 2 stroke:#8f7ac9
linkStyle 3 stroke:#8f7ac9
```
---
# Research mode
For questions that need sources rather than an opinion. Every claim in the answer must point to a page ZeroLog fetched, and the quote it relies on is checked against that page.
## Turn it on
- **`/research on`**: Switches the mode on in a chat. `/research off` switches it off. There is also a switch in the chat settings.
- **`/research `**: Not this mode: it starts a deep research run as a background job. See Deep research.
Chats with a container can switch to this mode by themselves for code and document questions. A sourced reply arrives in one piece instead of streaming.
## Check the sources
The answer comes with claims, each naming a source and an exact sentence from it. ZeroLog looks for that sentence word for word in the stored copy of the page. The result is shown per claim in the chat's **Sources** panel as supported, unsupported or uncited. If a check fails, the model gets one more try; if it still fails, the answer is shown with a warning.
The claim
A sentence quoted from a source
is looked up in
The stored page
Found word for word: supported.
*A quote that does not appear in the page is marked unsupported.*
If a page cannot be read, its source entry shows an error instead of a citation. Open the original from the panel to read the context.
## What follows
- **Check it instead of trusting it.** The check shows that a quote exists in the page, not that the page is right. Read what the claim rests on.
## What it costs
Slower and dearer than a normal answer, because a sourced reply can take several model calls. Use it when the citation matters.
[When you want a whole report **Deep research**](https://zerolog.ch/docs/deepresearch)
## How research mode checks a claim
Mermaid source of the diagram on this page:
```mermaid
%% Research mode: how a claim is checked. Facts from server/research.js (verifyClaim, GROUNDING_RULES, MIN_QUOTE), server/grounding.js (runGroundedTurn, one regeneration), web/apps/web/lib/slash.ts, components/chat/quellen-panel.tsx.
flowchart TB
on["/research on
or chat settings switch"]
fetch["Pages are fetched and
a copy is stored"]
ans["Model answers with claims,
each with a source and
an exact quote"]
chk["Is the quote in the stored
page, word for word?"]
ok["Supported"]:::hot
bad["Unsupported or uncited:
one more try, then
the answer carries a warning"]
panel["Sources panel lists
every claim and its verdict"]
on --> fetch --> ans --> chk
chk -->|yes| ok --> panel
chk -->|no| bad --> panel
linkStyle 3 stroke:#8f7ac9
linkStyle 4 stroke:#8f7ac9
```
---
Runs in your browser
# Voice
Speak instead of typing, and have replies read aloud. Recognition and speech use your browser's own APIs.
## Where the audio goes
Audio path
### Speaking
1. **You speak** In place of typing the message.
2. **Your browser handles it** Recognition is the browser's own API.
**Chrome, Edge**
Send the audio to Google to be recognised.
*Off your machine*
**Firefox, Safari**
No recognition at all. The interface says so.
*Path ends*
### Listening
1. **A reply arrives** The answer, as text.
2. **Your browser reads it aloud** Speech is the browser's own API too.
Dashed rail: the audio leaves your machine
*This server*
No audio reaches this server. There is no third-party voice key here, and nothing is billed for voice.
## What this does not mean
Audio never reaching this server is not the same as audio staying on your machine.
Chrome and Edge send it to Google to be recognised, which is the honest cost of this path. Firefox and Safari have no recognition at all, and the interface says so rather than failing quietly.
## Settings
Voice is switched on for new accounts, and an administrator can withdraw it from an account. Under Settings, Voice you choose how recording starts, either when you speak or while you hold the button, how long the pause before sending is (at least 800 ms), the recognition language such as `de-CH` or `en-US`, and whether replies are read aloud. Replies are spoken once they are complete.
---
Optional
# Buddy
A small collectible pet tied to your account. It appears as a card when you type `/buddy` in Claude Code inside a container terminal.
## Your pet
- **species**
- **rarity**
- **stats**
same look when you return
Species (one of 18), rarity from Common to Legendary, a possible shiny variant, eyes, a hat and five stats (debugging, patience, chaos, wisdom, snark) are worked out from your account. Nothing is stored for them, so your pet is always the same.
## Say hello
Type `/buddy` in Claude Code.
In a ZeroLog container, Claude Code prints your pet's card and greets you in its voice. Buddy is separate from the assistant and does not do your work.
## Name and limits
A name and personality are written once, by a low-cost model, when the pet hatches. Until then the card says the buddy is not hatched. The web chat has no Buddy corner and no `/buddy` command at the moment.
## Where Buddy comes from
Mermaid source of the diagram on this page:
```mermaid
%% Buddy: where the pet comes from and where it shows. Facts from server/buddy.js (buddyBones, species, rarities), server/routes/buddy.js (hatch), server/agent.js (provisionBuddyCommand).
flowchart TB
acc["Your account"]
bones["Species, rarity, shiny, eyes,
hat, five stats: worked out
from the account, never stored"]
soul["Name and personality:
written once at hatching
by a low-cost model"]
cc["/buddy in Claude Code
inside a container"]:::hot
card["Card in a code block,
greeting in the pet's voice"]
acc --> bones
acc -.-> soul
bones --> cc
soul -.-> cc
cc --> card
linkStyle 2 stroke:#8f7ac9
linkStyle 4 stroke:#8f7ac9
```
---
One step elsewhere, three here
# Import from ChatGPT and Claude
Bring your existing conversations in. The export comes from the other service, the rest happens here.
## The four steps
The other service
1. ### 01Export
The only step that happens over there. Request the data export in ChatGPT or Claude and open the archive you get back.
Here
1. ### 02Upload
In Settings, Privacy, Chat Import & Export, choose the file, drop it on the field or paste a single conversation.
2. ### 03Preview
The format is detected and you see what is inside before anything is written.
3. ### 04Confirm
You approve what the preview showed.
On large exports
Large imports run as a background job and show progress. The preview may not include a full conversation count.
## What it costs
Importing
Not billed
Continue one
Continuing
Costs credit
Importing does not use chat credits.
What lands is read-only history. It stays that way until you continue one of the imported chats, and continuing is the only part that costs credit.
## Which exports work
- **ChatGPT.** The `conversations.json` file from the ChatGPT data export. Very large files are read one conversation at a time.
- **Claude.ai.** The conversation file from the Claude data export.
- **ZeroLog.** A JSON file exported from ZeroLog can be imported again.
- **Claude Code.** Session files in `.jsonl` format. Sub-agent transcripts (`agent-*.jsonl`) are nested under the session that started them when both files come in together.
Upload the conversation file itself, not the whole archive. The size limit is your account's upload limit, never below 1 GiB. ZeroLog works out the format from the content, so you do not pick one. Importing the same file again updates the chats already there instead of duplicating them.
## What comes across
- **The transcript.** Text and code of every message, with the original timestamps where the export has them.
- **Attachments by name only.** An image or file shows up as a reference to its name, not as the file.
- **No tool or machine state.** Files an agent touched, tool state and anything kept on the provider side do not transfer.
- **Branches.** In a ChatGPT export, the line you ended on is imported. Discarded forks are counted and reported; tick “All branches” to bring them in as collapsed alternates.
Every import produces a stored report that lists what was imported, skipped or not recognised.
## Keep the context
Two ways to use old chats. To carry on one conversation, continue it: ZeroLog opens a new live chat that starts with the imported transcript. To draw on many old chats at once, tick “Make searchable” at import or later, and a new conversation can pull in relevant parts through the context index. The chat shows which sources it used. How that works is on [Context index](https://zerolog.ch/docs/stitch).
## Questions
### Does importing my ChatGPT history cost anything?
No. Importing does not use chat credit. Credit is used only when you continue an imported chat with a model, the same as in any other chat. The imported messages themselves are not charged again.
### Can I continue an imported ChatGPT or Claude conversation?
Yes. Choose continue on an imported chat and ZeroLog starts a new live chat from its transcript, so you can pick a model and keep going. Tool state and files from the original service do not carry over; the transcript does.
## From an export to a chat you can continue
Mermaid source of the diagram on this page:
```mermaid
%% Import: from another service's export to a chat you can continue. Facts from web/apps/web/components/settings/data-panel.tsx, server/imports.js (staging, worker job, dedupe, report), server/importers.js (detectFormat).
flowchart TB
exp["Export on the other service
(ChatGPT, Claude)"]
up["Upload or paste, check the
preview, confirm"]
job["Background job with progress,
stored report"]
hist["Read-only history
(not billed)"]
cont["Continue a chat:
new live chat, costs credit"]:::hot
srch["Make searchable:
context index"]
exp --> up --> job --> hist
hist --> cont
hist -.-> srch
linkStyle 4 stroke:#8f7ac9
```
---
Compute *→* Fast workspace
# Containers
Open Compute, choose Fast workspace, set a name and pick a system. The system is chosen when you create it. Free accounts have no container; every paid plan includes one.
## System
- Ubuntu 24.04 with ZeroLog toolchain
- Ubuntu 24.04, plain
- Debian 12 slim, Node.js 22
- Alpine 3.20
- Alpine, Python 3.12
Windows is not a container. It runs as a [virtual machine](https://zerolog.ch/docs/vms). Available systems and resource limits depend on your [plan](https://zerolog.ch/docs/plans).
## Build in your workspace
- `apt install`
- Compile
- Run a database
- Start a service
Install packages, compile code, run a database or start a service inside your workspace.
[Your isolated workspace](https://zerolog.ch/docs/isolation)
## It outlives the chat
Files persist across restarts and across chats. **The container is the durable thing; a chat is not.**
Attach a container to a chat so the AI works there. Attach the same container to several chats and they share one filesystem.
*Chat* *Chat* *Chat* *Restart* One filesystem
*Chats attach and end above it, the restart passes behind it, and the files stay where they are.*
A sketch of the relationship, not a timeline.
## Open in editor
Press **Open in editor** in the container overview or in the chat context bar. ZeroLog starts VS Code in your browser, which takes about 8 seconds, and opens it in a new tab.
The editor comes with Mini and higher. On Free the app points you to the [plans](https://zerolog.ch/docs/plans).
## Pause when idle
Pause a workspace while it is idle. Start it again when you need it. An unused workspace pauses itself after 30 minutes on Free, 1 hour on Mini, 4 hours on Pro, 8 hours on Max and 16 hours on Ultra.
*Kept* **Your files, on the server for 24 hours, then on network storage**
*Freed* **The cores and RAM**
## Snapshots
Save a known-good state before a risky change. Mini keeps 1 snapshot, Pro 2, Max 3, and Ultra up to 20 with daily automatic ones.
Known-good state
The risky change
## Resize, rebuild, export
Under **Advanced** you can change cores, RAM and disk within your plan. **Rebuild** resets the runtime to the base image and keeps your code in /work; a snapshot is taken first and packages installed outside /work are lost. **Export** produces a ZIP with a Dockerfile, a compose file and your source.
## Coding agents
The Agents page lists the coding agents in your workspaces. Set one up, start a session, sign in to the provider, update it or check what is wrong, per workspace. A session can be opened in your local terminal.
## What a plan gives one container
Mini 1 vCPU, 4 GB RAM, 10 GB disk. Pro 2 vCPU, 8 GB, 15 GB. Max 4 vCPU, 12 GB, 25 GB. Ultra runs up to three containers that share 12 vCPU, 48 GB RAM and 150 GB disk. If a plan change leaves you over quota, the workspaces that no longer fit are paused, never deleted.
## How the AI finds its way
Besides reading single files, the AI can search the whole workspace with a regular expression: matching lines with their line numbers and optional surrounding lines, only the names of matching files, or the number of matches per file. It can limit the search to a folder or to file names such as `*.ts`. The search skips `.git`, `node_modules` and binary files, returns at most 500 results and says how many more there are. It only reads; it can read what the AI could open as a file anyway.
It can also list files by pattern, such as `src/**/*.ts`, newest first with their modification time, at most 1000 at a time.
When it reads a PDF, it gets the text page by page, up to 20 pages at a time; the text is extracted inside your workspace with poppler, and the AI tells you if poppler is missing. A Jupyter notebook comes back as its cells with their source and outputs; images in outputs are only named. To change a notebook, the AI inserts, replaces or deletes single cells instead of rewriting the file; a replaced code cell loses its old outputs.
What each plan includes, with current prices: [Pricing](https://zerolog.ch/pricing).
## Life of a workspace
Mermaid source of the diagram on this page:
```mermaid
%% Containers: life of a workspace. Facts from server/containers.js (create, pause, resume, enforceQuota, destroy), osimages.js, tiers.js, web/apps/web/components/compute/.
flowchart LR
create["1. Create
Compute, Fast workspace
name and system"]
run["2. Running
files persist
chats attach"]
paused["Paused
by you, or over quota
after a plan downgrade
files on network storage
cores and RAM freed"]:::hot
snap["Snapshot
saved state
of the files"]
rebuild["Rebuild
snapshot first
runtime reset
/work is kept"]
gone["Deleted
you type the name
to confirm"]
create --> run
run -->|Pause| paused
paused -->|Start| run
run -->|Snapshot| snap
snap -->|Restore| run
run -->|Rebuild| rebuild
rebuild --> run
run -->|Delete| gone
classDef hot stroke:#8f7ac9
```
---
# Container start
A real percentage while a container starts, and "Running" only when it is ready to use. Both come from measured steps, not from a timer.
- **The steps:** planning the start, bringing the workspace back from network storage (only if it was moved there), checking the image, preparing the network, mounting the workspace, creating the machine, booting it, starting the inner sandbox (only for the VM with gVisor class), checking network access, and a readiness check. The bar shows the step you are in and its percentage.
- **Where the number comes from:** every finished step counts with its weight. The step in progress adds a part of its own weight: bytes unpacked while the workspace comes back, probe attempts during the readiness check, or elapsed time against the usual time for that step. The number never goes past the weight of the current step and never reaches 100 before the container runs.
- **Taking longer than usual:** if a step takes three times its usual time, the number stops and the view says so. It does not creep forward to look busy.
- **Running means ready:** a command runs inside the machine, the sandbox network has a gateway (so the terminal can connect), and, for the VM with gVisor class, the inner sandbox answers. The file list is tried with a short deadline. Only after that does the container show **Running** and 100 %. Until then the terminal answers with "still starting" instead of an error, and the Files view opens when the container runs.
- **If a step fails:** the view names the step and the reason, the container goes back to the state it was in, and **Try again** starts it again. A machine that booted but failed the readiness check is stopped, so it never shows Running next to an error.
- **The agent runtime** (used by Tailscale and by agents in your chats) is its own phase after the start. It is not a condition for Running; opening Tailscale or the terminal waits for the same start instead of building a second one.
- **Measured, not guessed:** each step's duration is stored. The weights start from one cold-start measurement and are pulled from the medians as data comes in.
- **For API clients:** send the header `X-ZL-Start: async` with `POST /api/containers/:id/resume` to get `202` and poll `GET /api/containers/:id/wecken`. The `wecken` object carries `phase`, `phase_label`, `prozent`, `schritt`, `fehler` and `fehler_phase`. Without the header, the call waits and answers `200` as before, and `wecken` is included.
Pausing and storage are described under [Containers](https://zerolog.ch/docs/containers). The Tailscale connection is under [Tailscale in a container](https://zerolog.ch/docs/tailnet).
## From Start to Running
Mermaid source of the diagram on this page:
```mermaid
%% Container start: from the click to Running. Facts from server/containerstart.js (phases, weights), server/sandbox.js (starteLaufzeit, bereitschaft), server/containerzustand.js.
flowchart TB
a["1. You press Start
state: Starting
0 %"]
b["2. Plan, image,
network, workspace
real steps, measured"]
c["3. Machine created
and booted
VM boot is the longest step"]
d["4. Readiness check
command runs, network gateway,
inner sandbox, file list"]:::hot
e["5. Running
100 % only now
terminal and files open"]
f["A step fails:
step and reason shown,
Try again"]
g["Step 3x slower than usual:
number stops,
view says so"]
a --> b --> c --> d --> e
b -->|error| f
c -->|error| f
d -->|error, machine stopped| f
c --> g
classDef hot stroke:#8f7ac9
```
---
# Isolation
Your workspace is its own container, not shared with any other user.
- **Your own workspace**
Nobody else's workspace can read your files, see your processes or reach your running services.
- **An extra layer under the container**
ZeroLog runs workspaces on [gVisor](https://gvisor.dev), a userspace kernel: your workspace's system calls are handled without reaching the host kernel directly. A machine without gVisor does not start workspaces at all.
- **No shared processor cores**
The machines that run workspaces never execute at the same moment as other machines on the two halves of one physical processor core, and identical memory pages are not merged with other machines. Both close known ways for one machine to observe another through shared hardware.
- **The layer under your workspace is itself fenced in**
The host process that runs your workspace cannot read ZeroLog's own keys, database or configuration, cannot reach the container engine, and cannot change the rules that fence it in. If that layer were ever broken out of, the next wall is already standing.
- **No path between workspaces**
Workspaces cannot reach each other over the network, even when they run on the same machine.
- **Outbound traffic is filtered**
A workspace's own internet access goes through ZeroLog's network protections; see [Outbound network](https://zerolog.ch/docs/egress).
## Choose a stronger class for risky code
When you create a container you can pick its isolation class. **Standard** runs the workspace on gVisor with an extra system call filter that blocks interfaces a development environment does not need. **VM** gives the workspace its own small virtual machine with its own kernel, inside which it runs; use it for untrusted code. A VM workspace starts a few seconds slower and cannot be switched back later.
The VM class is part of the Ultra plan. Mini, Pro and Max run the standard class, and Free has no workspace. See [Plans](https://zerolog.ch/docs/plans) and [Pricing](https://zerolog.ch/pricing).
## Previews cannot reach your account
Websites, apps and designs you preview run sealed off from ZeroLog itself: they cannot read your session, your chats or anything else in your account, even when they come from code the AI wrote.
## What this does not change
### Commands still wait for you
Isolation keeps workspaces apart from each other, not from you. A command the AI wants to run still waits for your approval first — see [Command approval](https://zerolog.ch/docs/approval).
### Secrets go through the vault
Credentials you attach are handled by the vault, not pasted in the clear — see [Secrets vault](https://zerolog.ch/docs/vault).
### Pausing keeps the disk
Pausing a workspace frees its running slot; it does not delete anything. A paused workspace's disk stays until you delete it yourself.
Credentials for agents: [API key proxy for AI agents](https://zerolog.ch/api-key-proxy-for-ai-agents).
## Where a workspace runs and how traffic leaves it
Mermaid source of the diagram on this page:
```mermaid
%% Isolation: where a workspace runs and its only way out. Facts from server/sandbox.js, server/seccomp.js, server/broker.js, keyservice/.
flowchart TB
user["Browser, CLI or app"]
app["App server (control plane)
sign-in, chats, approvals"]
dfilter["Filtered container API
no Docker socket
in workspaces"]
subgraph WK["Worker machines, one workspace each"]
direction LR
subgraph STD["Class: standard"]
gv["Your workspace
gVisor user-space kernel
all capabilities dropped
internal network only"]:::hot
end
subgraph VM["Class: vm (Ultra plan)"]
vm["Your workspace
Kata micro-VM
own guest kernel
internal network only"]
end
end
broker["Egress broker
runs inside the app process
target and port rules"]
net["Public internet"]
subgraph KS["Key service (separate VM)"]
ks["Third-party keys, encrypted
mutual TLS from app host only"]
end
api["Third-party APIs"]
user --> app
app -->|create, stop| dfilter
dfilter -->|start, stop| WK
gv -->|only way out| broker
vm -->|only way out| broker
broker -->|allowed targets| net
broker -->|hosts with a key profile| ks
ks -->|real key inserted| api
classDef hot stroke:#8f7ac9
```
---
# Command approval
Every command the AI wants to run appears in full and waits. Three buttons end that wait, and the middle one changes more than this one call.
## What you see when it stops
waits here
the command, in full
When the command can be edited, that box is a text field. Change the text, press Approve, and what runs is what you typed.
Approve Always in this chat Deny
A line here says what **Always in this chat** does: it ends the questions for the whole chat. A chat has four settings: Ask every time, Allowlisted commands only, Auto-accept edits and Never ask.
- **Approve**: Runs it once.
- **does more than this callAlways in this chat**: Switches the whole chat to Auto-accept edits, so every later tool runs without asking. Change it back under the chat's settings.
- **Deny**: Refuses and returns `[denied by user]` to the model.
*Deny is not next to Approve, so a destructive command is never one mis-tap away.*
fewer questions
## The command allowlist
Set a chat to **Allowlisted commands only** and commands on your list run without asking. Everything else still waits for you. When a command asks, **Always allow** approves it and adds a pattern to the list. Edit the list under Settings, Command allowlist; it applies to all your chats.
how a pattern matches
- `npm test` allows exactly that command
- `npm run:*` allows that start with any arguments
- in a chained command (`;` `&&` `||` `|`), every part must be on the list
- `$(…)`, backticks, redirection and `\` always ask
if you do not respond
## The request expires
After five minutes, the pending request expires. The assistant stops and reports what remains undone.
after it runs
## In the chat
A command that runs keeps its exit code and output in the chat.
a command card holds
- the command
- its exit code
- its output
## What happens to one command
Mermaid source of the diagram on this page:
```mermaid
%% Command approval: what happens to one command the AI wants to run. Facts from server/routes/chats.js (gateFor, approval wait, 5 min), server/befehlsliste.js, web/apps/web/components/chat/tool-card.tsx.
flowchart TB
ask["1. The AI asks to run a command"]
mode["2. The chat's setting decides"]
wait["3. Shown in full,
waits for you"]:::hot
run["4. The command runs
exit code and output
stay in the chat"]
deny["Denied
the model gets
[denied by user]"]
late["Request expires
the assistant stops
and sums up"]
ask --> mode
mode -->|"Never ask, Auto-accept edits"| run
mode -->|"Allowlisted: every part is on your list"| run
mode -->|"Ask every time, or not on the list"| wait
wait -->|"Approve, Always in this chat, Always allow"| run
wait -->|Deny| deny
wait -->|"no answer in 5 min"| late
classDef hot stroke:#8f7ac9
```
---
# Chat permissions
Every chat has two separate permission settings, both under the sliders icon in the chat: one for commands and file changes in containers, and one for tools the server runs itself. You can change either while the AI is working.
A sorting machine for tool calls: a belt, four stations that hang down depending on a lever, a ram that pushes flagged calls to a side lane, and a gate with a button.
*A sorting machine for server tool calls. The lever picks which stations hang down; flagged calls wait on the side lane for the button. A call's height is its price, and the price bar hangs at 0.50 USD.*
Figure as text
1. Run without asking. No station hangs down, so every call rides straight to the output, even one that costs money.
2. Ask first. A feeler hangs just above free calls. Any call that costs money touches it, is flagged and is pushed to the side lane, where it waits at a gate for your button.
3. Automatic review. Fixed rules hang first: a call that reaches outside the platform, or is estimated above 0.50 USD, is flagged. A free call with no outside effect passes. For the rest, a gauge stands in for the reviewer model: it has 3 seconds, and a doubtful or slow answer sends the call to the side lane.
4. Manual. A blocker hangs low enough to flag every call. The calls wait on the side lane as outlines, only proposed. Start turns one into a real call and lets it run.
5. The lever can move while calls wait. Run without asking opens the gate for them; Manual turns them into outlines.
## Container commands
Four modes: **Ask every time**, **Allowlisted commands only**, **Auto-accept edits** and **Never ask**. **Never ask** needs a second click (**Understood, switch**). The prompt and the allowlist are described under [command approval](https://zerolog.ch/docs/approval).
## Server tools
The **Server tools** setting covers jobs the server runs, such as scans, designs and sites, plus the paid helpers for asking another model and for images.
- **Run without asking:** runs at once, even if it costs money.
- **Ask first:** an approval card for every call that costs money. Free calls run.
- **Automatic review:** fixed rules decide first. The chat asks you when a call reaches outside the platform, cannot be undone (delete, publish, deploy, send, pay), points at someone else's address, has an unknown price, or is estimated above 0.50 USD. A free tool with no outside effect runs. For everything else a small reviewer model, the one that also summarises long chats, sees the redacted call and answers allow or ask within 3 seconds. An error, a timeout, an unclear answer or no reviewer available (no balance, anonymous chat, zero-data-retention chat) means it asks you. The card says who decided and why. Reviewer calls are billed to your credit like other internal calls.
- **Manual:** the AI only proposes. Nothing runs, and the AI is told so and carries on. The chat shows a **Start** button; press it and the result replaces the proposal in the same row. Start works once, only for the chat's owner, and needs credit. The paid helpers cannot be started later and behave like Ask first.
- A chat with no choice of its own follows its container mode: Auto-accept edits or Never ask gives Run without asking, anything else gives Ask first.
## Switching mid-turn
The level is read again for every tool call. Switch the container mode to Auto-accept edits and approvals already waiting in the running turn are granted at once; later calls run unasked, and switching back asks again. For server tools, Run without asking grants the waiting cards and Manual turns them into proposals. While the AI waits in ask mode, a hint offers **Switch to Automatic?**. Only the owner of the chat can switch.
## How Automatic review decides one call
Mermaid source of the diagram on this page:
```mermaid
%% Chat permissions: how Automatic review decides one server-tool call. Facts from server/werkzeugstufe.js and server/routes/chats.js.
flowchart TB
a["1. The AI calls a server tool
level is read at that moment"]
r["2. Fixed rules first"]:::hot
h["You decide
approve or deny"]
run["The tool runs"]
rev["3. Reviewer model
allow or ask within 3 s"]
a --> r
r -->|outside effect, no undo,
unknown or high price| h
r -->|free and harmless| run
r -->|undecided| rev
rev -->|allow| run
rev -->|ask, error, timeout| h
h -->|approve| run
classDef hot stroke:#8f7ac9
```
---
# Files
A real file tree of the workspace. Open **Compute**, then **Files**, and pick a workspace.
## What you can do
- browse
- open a file, rendered or as source
- edit a text file and save it
- download a single file
A paused workspace keeps its files on the server. Once they have moved to network storage (after 24 hours paused), Files shows the last listing. Restore the workspace to open or download files.
## Getting files in
- attach them in a chat
- mount the workspace as a [network drive](https://zerolog.ch/docs/smb)
- use the editor or the [terminal](https://zerolog.ch/docs/terminal)
## What the AI sees
Edits you make are immediately what the AI sees. If the file changed after you opened it, your save is refused and you reload first.
you edit a file
the AI sees it
the same filesystem
no sync step
## Browsing and editing workspace files
Mermaid source of the diagram on this page:
```mermaid
%% Files: one workspace's file tree. Facts from web/apps/web/components/compute/files.tsx, components/chat/file-viewer.tsx, server/routes/containers.js (files-metadata, files, file, text, POST file).
flowchart TB
tab["1. Compute, Files
pick a workspace"]
live["Running workspace
live file tree"]
off["Paused workspace
last listing from
network storage"]
restore["Restore workspace"]
open["2. Open a file
rendered or as source"]
save["Edit a text file and save
refused if the file changed
since you opened it"]:::hot
dl["Download one file"]
ai["Same filesystem
the AI sees the change"]
tab --> live
tab --> off
off --> restore --> live
live --> open
live --> dl
open --> save --> ai
classDef hot stroke:#8f7ac9
```
---
# Terminal
A shell in the same container, available mid-task without stopping the agent. It needs a plan with a container, Mini or higher.
## Stepping in mid-task
Same container
Agent You Open Hand back
*The agent keeps working while you are in the shell. **Fix one thing by hand and hand it back**; the conversation keeps its context.*
## A full terminal
**htop**, **vim** and **tmux** behave normally.
## Eleven coding assistants
A picker in the terminal panel lists eleven entries. It groups them by ZeroLog models and your own account, and shows each one's model, setup state and approval.
- Claude Code, and Claude Code with your own account
- Codex
- ZeroLog Code
- opencode
- Kilo Code
- goose
- pi
- Qwen Code
- Nanocoder
- Kimi Code
Claude Code and Codex are part of the container image. The others install the first time you open them, from pinned versions with checksums and only inside your own container. After a container rebuild they come back from a cache within seconds.
Nine of them run on ZeroLog models and use your normal **credit**. Claude Code talks to an Anthropic-style endpoint (`/v1/messages`); the others use the OpenAI-style endpoint `/v1/chat/completions`. Codex and Claude Code with your own account use that account instead and spend no ZeroLog credit. Cancelling a request stops it at the provider, and a stream is billed once even when you cancel. The picker has a form to suggest another tool.
## When context cannot be carried over
When you open a coding CLI from a chat, ZeroLog tries to carry over the chat context. If the output is too large or the transfer takes too long, it skips the transfer. The terminal still opens.
If ZeroLog cannot read your existing project instructions, it leaves them unchanged. If the CLI is missing context, give it the relevant instructions directly.
## How an assistant gets into your container
Mermaid source of the diagram on this page:
```mermaid
%% Terminal: how a coding assistant gets into your container and where its model calls go. Facts from server/terminal-clis.js, server/terminal.js, server/anthropic-shim.js.
flowchart TB
pick["1. Open the terminal panel
and pick an assistant"]
img["Claude Code, Codex
already in the container image"]
fetch["2. First open of any other one:
the host fetches the pinned version
and checks its checksum"]
unpack["3. Unpacked into your own container only
checksum checked again there"]:::hot
start["The assistant starts"]
zl["ZeroLog models
through the gateway
your normal credit"]
own["Codex, Claude Code
with your own account
no ZeroLog credit"]
pick --> img --> start
pick --> fetch --> unpack --> start
start --> zl
start --> own
classDef hot stroke:#8f7ac9
```
---
# Network drives
Available
Open workspace files from your computer with WebDAV or SMB.
---
From anywhere
## Use WebDAV
Open the Storage tab of a workspace. It shows the address and setup command for your computer. WebDAV works over HTTPS on the same address as the site, with no extra port and no VPN, so you can use it away from the ZeroLog management network.
- 1 · Address
### Copy it from the panel
Use the address shown for your workspace.
- 2 · Sign in
### Use your file access details
The panel shows your username. The password is the one you set there, and it serves both WebDAV and SMB.
- 3 · Connect
### Follow your system's command
Windows (`net use` in PowerShell), macOS and GNOME (the address) and Linux (davfs2) are shown in the panel. On Android, pair the ZeroLog app with the QR code. On iPhone and iPad, use an app that supports WebDAV.
---
For local networks
## SMB is limited to the management network
Only admins can see and switch on an SMB share in the Storage tab. SMB only works inside the management network and may not be available for every workspace. For remote access, use WebDAV. On Windows, [ZeroLog Connect](https://zerolog.ch/docs/connect) can also mount your own container as a drive through an HTTPS tunnel.
A new file access password replaces the old one. The password is shown once when you set it.
---
Scope
## Workspace files
Network drives include workspace files. Virtual machine disks are not available as network drives.
## Two ways to mount a workspace as a drive
Mermaid source of the diagram on this page:
```mermaid
%% Network drives: two ways to mount a workspace. Facts from server/dav.js, server/smb.js, web/apps/web/components/containers/storage-panel.tsx.
flowchart TB
start["1. Storage tab of your workspace"]
pw["2. Set the file access password
shown once, a new one
replaces the old one"]
dav["WebDAV over HTTPS
same address as the site
works from anywhere"]:::hot
smb["SMB, after you switch
a share on
management network only"]
desk["Windows: net use in PowerShell
macOS, GNOME: the address
Linux: davfs2"]
phone["Android: ZeroLog app, paired by QR
iPhone, iPad: a WebDAV app"]
drive["3. Workspace files as a drive
read and write
machine disks are not included"]
start --> pw
pw --> dav
pw --> smb
dav --> desk --> drive
dav --> phone --> drive
smb --> drive
classDef hot stroke:#8f7ac9
```
---
# Drive and SSH in one click
Mount your own container as a drive and open SSH to it from Windows without typing a command. You press a button on the container's **Access** tab, and a small helper, ZeroLog-Connect.exe, does the rest over HTTPS. Only the owner of the container can use it.
A pneumatic tube from your PC to your own container. A button makes a one-time code, a helper turns it into a device token, and a capsule carries it through the tube to an airlock that opens for the owner only.
*A tube from your PC to your container. Only the owner passes the airlock. The one-time code lasts 5 minutes and works once. The device token lasts 90 days from last use.*
Figure as text
1. Press Connect drive or Open SSH on the container's Access tab. The browser makes a one-time code that works once, for 5 minutes, and hands it to the helper through a zerolog-connect:// link, never in a web address.
2. The helper swaps the code for a device token, valid 90 days from last use. The token opens the tunnel for that container only.
3. The token rides through an HTTPS tunnel that upgrades to a WebSocket. No extra port opens on your PC.
4. At the container end a reader checks the account. Only the owner passes. A member of a shared container, or an admin who does not own it, gets a refusal and the door stays shut.
5. For the owner the door lifts. The drive is an SMB share served on the same machine's loopback address and carried through the tunnel. SSH goes to the container through its gateway.
6. Disconnect closes the door at once, including open connections. A locked account stops working immediately.
- **Windows first.** On macOS and Linux the commands on the same tab still apply. The drive needs Windows 11 24H2 or later.
- **First time:** on the Access tab, under **Connect in one click**, download ZeroLog-Connect.exe and start it once. It needs no administrator rights. It keeps its credentials in Windows Credential Manager and stays in the notification area while a drive is connected.
- **The file is not code-signed yet,** so Windows SmartScreen warns when you start it: choose **More info**, then **Run anyway**. The page shows the file's SHA-256. Compare it first with `Get-FileHash .\ZeroLog-Connect.exe` in PowerShell. Everyone gets the same file, with no account or code inside it.
- **Connect drive** shows the container as a network drive, from the letter Z downward. The drive is an SMB share carried through an HTTPS tunnel; no extra port opens on your side. The share is switched on at your first click. This is separate from [WebDAV](https://zerolog.ch/docs/smb).
- **Open SSH** sets up a key, a pinned host key and an entry in `~/.ssh/config`. Afterwards `ssh zl-NAME` works in any console and in VS Code Remote-SSH. There is no "trust this host?" prompt.
- **How it is secured:** a click creates a one-time code that works for 5 minutes and once. The browser hands it to the helper through a `zerolog-connect://` link, never in a web address. The helper swaps it for a device token valid 90 days from last use. The token only opens the tunnel for that container.
- **Connected devices** lists every PC you linked. **Disconnect** ends one at once, including open connections. **End all connections** also resets the file password, so WebDAV and drives need reconnecting.
- Members of a shared container and admins who do not own it get a refusal. A locked account stops working immediately.
Status: this is released, but a complete run on a customer's Windows PC has not been confirmed yet. If the drive or SSH does not connect, tell us what the helper says. See also [terminal](https://zerolog.ch/docs/terminal) and [files](https://zerolog.ch/docs/files).
## From the button to the drive or SSH
Mermaid source of the diagram on this page:
```mermaid
%% Connect in one click: from the button to the drive or SSH. Facts from server/connectklick.js, server/routes/connect-klick.js, web/apps/web/components/containers/connect-klick.tsx.
flowchart TB
a["1. Access tab
Connect drive
or Open SSH
owner only"]
b["2. One-time code
5 minutes, once
zerolog-connect:// link"]
c["3. ZeroLog-Connect.exe
swaps the code for a
device token, 90 days"]:::hot
d["4a. Drive
SMB through an HTTPS
tunnel, letter from Z"]
e["4b. SSH
key, pinned host key,
ssh zl-NAME"]
f["Disconnect one device
or end all connections
takes effect at once"]
x["First time:
download the helper
SmartScreen warns
compare the SHA-256"]
x --> a
a --> b --> c
c --> d
c --> e
d --> f
e --> f
classDef hot stroke:#8f7ac9
```
---
# Public ports
Start something listening in the container, then publish that port from the workspace's Network tab.
## What publishing does
Inside the container
Something you started is listening on a port, bound to `0.0.0.0`. A service bound only to `127.0.0.1` cannot be published.
Publish
Public
https:// \* .zerolog.ch
\* is name-account: your account name is part of the hostname
The URL works immediately, certificate included.
Because your account name sits in the hostname, nobody can squat yours.
## How many you get
How many ports you may publish is set by your [plan](https://zerolog.ch/docs/plans): Free none, Mini 1, Pro 1, Max 2, Ultra 8. If the limit is used up, withdraw a port first. A withdrawn address stops working at once.
## The first publish
The first publish can need one rebuild of the workspace. The panel names the running sessions that will end and asks before it restarts.
## Anything published is public.
Put authentication in front of anything that should not be.
## From a listening service to a public URL
Mermaid source of the diagram on this page:
```mermaid
%% Public ports: from a listening service to a public URL. Facts from server/ingress.js (publish, hostnameFor), server/tiers.js (public_ports), web/apps/web/components/containers/ports-panel.tsx.
flowchart TB
a["1. A service in your workspace
listens on 0.0.0.0
not only on 127.0.0.1"]
b["2. Network tab
the panel lists what
is listening now"]
c["3. Pick a name and the port
and publish"]
q["Quota used up:
withdraw a port first"]
r["First publish can need one
rebuild of the workspace
running sessions end"]
url["4. Public HTTPS URL
name-account.zerolog.ch
certificate included
open to anyone who has it"]:::hot
w["Withdraw
the address stops
working at once"]
a --> b --> c
c -->|over your plan's limit| q
c --> r
c --> url
url --> w
classDef hot stroke:#8f7ac9
```
---
# Preview links
When the AI starts a web app in your container, the chat shows it in a side panel. Its address is a preview link. The link carries its own permission, so it opens without a ZeroLog sign-in, and for that reason it expires.
- **Expiry:** a preview link lasts 7 days by default. When you create one yourself you can choose 1 hour, 24 hours, 7 days or 30 days. Nothing lasts longer than 30 days.
- **Where:** the **Links** bar above the message box lists what is open in the chat. A green dot means something is listening on that port right now. Open **Sharing** to see when each link expires.
- **Share:** pick **Valid for**, then **Create share link**. Treat the address like a password: anyone who holds it can open the page until it expires or you revoke it.
- **Revoke:** **Revoke** ends one link; **Revoke all links** ends every link of the container, including older ones. Whoever opens a dead link sees "Preview no longer available" (HTTP 410), and you can create a new one.
- **Automatic links** for the side panel are reused while they have at least an hour left. Otherwise a fresh 7-day link is made.
- Only the owner can create and revoke links. Members of a [shared container](https://zerolog.ch/docs/share) cannot.
Previews run sealed off from your account, see [isolation](https://zerolog.ch/docs/isolation). They are not [public ports](https://zerolog.ch/docs/ports): they use no port quota and no permanent address.
---
# Custom domains
On Pro, Max and Ultra you can point your own domain at a [published port](https://zerolog.ch/docs/ports).
- Pro **One domain**
- Max **One domain**
- Ultra **Unlimited**
Yours
- Your domain name
- Your registrar, where the record goes
ZeroLog
- Routing to the published port
- Issuing the certificate
***DNS** Add the records shown in your account, in the order provided: TXT, then CNAME.*
## Set up your domain
Under **Custom domains** in Compute, Connections and drives, add your domain and choose a published port, or assign one later. Add the TXT record first: it proves you control the domain and gets the certificate issued. Then add the CNAME to the target the panel shows. Check the domain status in the panel; it reads live when both records are in place.
## Removing a domain
**On removal** Remove the domain, and both the certificate and the routing are cleaned up.
## Pointing your own domain at ZeroLog
Mermaid source of the diagram on this page:
```mermaid
%% Custom domains: DNS order. Facts from server/ingress.js (addDomain, domainStatus, removeDomain), web/apps/web/components/containers/account-cards.tsx.
flowchart TB
add["1. Add your domain
Connections and drives,
Custom domains
port now or later"]
txt["2. First DNS record: TXT
proves you control the domain
the certificate is issued"]:::hot
cname["3. Second DNS record: CNAME
to the target in the panel
once the TXT is valid"]
live["Status: live
check it in the panel"]
gone["Remove the domain
it stops pointing here
certificate and routing go"]
add --> txt --> cname --> live
live --> gone
classDef hot stroke:#8f7ac9
```
---
# Internet access
Container connections are protected by ZeroLog's network safeguards.
## What you can expect
Most common software sources and websites work without extra setup.
1. Usually
### Common destinations
Package sources and websites your work depends on.
Available
2. Sometimes
### A destination may be unavailable
Some connections are restricted to reduce safety and abuse risks. Internal and private addresses are never reachable from a workspace.
Restricted
3. Always
### A few service ports stay closed
Mail delivery, Windows file sharing and a handful of old services that outbound traffic only ever uses for spam or password guessing.
**No setting opens them**Not a permission level, not an allow-list entry.
Closed
4. Watched
### Unusual patterns slow down
Many different destinations in a short time, many ports on one host, or the same service tried over and over.
**Ordinary work is untouched**Installing packages, cloning a repository, pushing a large commit.
Throttled
5. Error
### Check the command output
A failed connection can also come from the destination or a temporary network problem.
Retry later
### Still cannot connect?
Include the destination and the error message when you contact ZeroLog support.
Get help
## Testing your own servers
Scanning is allowed inside a scope you have proven you own: verify the domain, and the scan may reach it and its subdomains. A load test needs a separate, time-boxed allowance — from the outside it looks exactly like an attack, so it is granted per target and expires on its own. Ask support with the target and the window you need.
## Container credentials
Each container reaches the internet with a network credential that belongs to that one container. A credential presented from another container, from an unknown container type, or for a deleted or offloaded container is rejected. Support can rotate the credential of a single workspace: the old one stops working at once and the workspace gets the new one when its runtime is recreated. One limit remains: containers on worker nodes reach the gateway through a tunnel that does not reveal their name, so this name check does not apply to them yet.
## How one outbound connection is judged
Mermaid source of the diagram on this page:
```mermaid
%% Outbound network: how one connection from a container is judged. Facts from server/broker.js, server/egress.js (checkAcl), server/abwehr/portpolitik.js, server/abwehr/netzwaechter.js.
flowchart TB
c["Your container
no route to the internet
except through the broker"]
b["Egress broker
knows which container
and which owner"]
p["1. Is the port closed?
mail, SMB, telnet and other
old services: refused, always"]
g["2. Is the destination
on your allow-list?"]
d["3. A public host that is on
no blocklist? Allowed by default.
Internal addresses stay refused"]:::hot
net["Internet"]
no["Refused and recorded"]
c --> b --> p
p -->|closed| no
p -->|open| g
g -->|yes| net
g -->|no| d
d -->|"yes (slowed if the pattern is unusual)"| net
d -->|no| no
classDef hot stroke:#8f7ac9
```
---
# Tailscale in a container
Add a container to your own Tailscale network with one click, so the AI in your chats can reach machines inside it, such as a home server or a database behind your firewall. You sign in at Tailscale. ZeroLog never sees your password and never stores a Tailscale auth key for this.
A rack with your container joined to your own Tailscale network. Its node is ephemeral and shielded. An outbound pipe runs through an egress gate to a relay mast, a read-only status board lists your devices, and a SOCKS jack lets the AI reach them.
*Your container joins your tailnet for up to 4 hours. It reaches your devices through a local proxy. Nothing reaches in.*
Figure as text
1. Press Connect to Tailscale. A new tab shows the Tailscale sign-in. You sign in there; ZeroLog never sees your password and stores no auth key. You have 10 minutes.
2. While the sign-in is open, the egress gate lets the container's node reach Tailscale. The gate opens only for a live session and shuts again when it ends.
3. The node runs in userspace under its own user, with shields up, no subnet routes and no exit node. It is ephemeral and lasts 4 hours by default.
4. A read-only status board lists your devices: name, system, online state and addresses. It holds no login link and no keys.
5. The AI reaches a device through the local SOCKS5 proxy at 127.0.0.1:1055. An offline device does not answer.
6. Nothing comes in. Your devices cannot reach the container. Disconnect signs the node out and the gate shuts.
- **Where:** open a running container and find the **Tailnet** section in its overview. You need to be the owner, have a plan with containers (Mini or higher) and the Tailscale permission, which new accounts have.
- **Connect:** press **Connect to Tailscale**. A new tab opens and shows the Tailscale sign-in. If the container is paused, it wakes first. If the work environment is still starting, the view says so (about 2 minutes on a cold start), then "Tailscale starts" and "Fetching the sign-in page"; the tab fills itself in. Sign in there and the view notices on its own, then shows the device name, your tailnet and its 100.x address. **Cancel** ends the attempt while it waits; the work environment keeps starting, so the next try is quicker. A tab that gets no page tells you why instead of closing. After 3 minutes without a sign-in page the view stops waiting and offers **Try again**. If the container cannot reach Tailscale at all, you get that message after one minute. Opening the section already warms the work environment.
- **Time limits:** finish the sign-in within 10 minutes or the attempt ends. A connection lasts 4 hours by default, and the view shows when it ends. It also ends when the container pauses or restarts. Connect again afterwards. You can start 10 attempts per 10 minutes.
- **Disconnect** signs the device out of Tailscale and stops it. If Tailscale does not confirm the sign-out, the device drops out of your tailnet shortly on its own.
- **One direction only.** The container reaches your tailnet; your devices cannot reach the container (shields-up). It runs in userspace networking, with no subnet routes and no exit node. The device is ephemeral: it disappears from your tailnet when the connection ends. Its name comes from the container name.
- **What the AI sees:** a read-only list of your devices (name, system, online state, addresses) in `/tmp/zl-tailnet-status.json`, refreshed by ZeroLog. It reaches a device through the local SOCKS5 proxy at `127.0.0.1:1055`, for example `curl --socks5-hostname 127.0.0.1:1055 http://name:port`. The list holds no login link and no keys.
- **No Tailscale command for the agent.** The `tailscale` CLI is not available to the AI by design, so a failing `tailscale status` does not mean the connection is off.
This is separate from the [key proxy](https://zerolog.ch/docs/keyproxy), which can hold a Tailscale API token, and from [desktop control](https://zerolog.ch/docs/desktop), which works on your own computer. Container traffic otherwise follows [outbound network](https://zerolog.ch/docs/egress) rules. This describes the released behaviour; a full sign-in on a customer account has not been recorded yet, so tell us if it fails.
## Connecting a container to your tailnet
Mermaid source of the diagram on this page:
```mermaid
%% Tailnet in a container: from the click to the AI reaching your devices. Facts from server/tailnet-runtime.js, server/tailscale.js, web/apps/web/components/containers/tailnet-panel.tsx.
flowchart TB
a["1. Container overview
Tailnet section
Connect to Tailscale"]
b["2. New tab opens
Tailscale sign-in
up to 3 min on a cold start"]
c["3. You sign in at Tailscale
ZeroLog never sees
your password"]
d["4. Connected
ephemeral device
userspace, shields-up
no inbound"]:::hot
e["5. The AI reads
/tmp/zl-tailnet-status.json
and uses SOCKS5
127.0.0.1:1055"]
f["Ends: Disconnect,
pause, restart
or after 4 h"]
x["Sign-in not finished
after 10 min:
attempt ends"]
a --> b --> c
c --> d
c -->|too slow| x
d --> e
d --> f
classDef hot stroke:#8f7ac9
```
---
# Virtual machines
Some work does not fit in a [container](https://zerolog.ch/docs/containers): a different kernel, systemd, Windows. For those cases you can create a full virtual machine from a template an admin has defined, within your quota. Virtual machines come with the Ultra plan.
## Where a container is not enough
*Plan view*
Container
- a different kernel
- systemd
- Windows
Virtual machine
## The bounds
- **Template**: Choose from the templates listed for your account.
- **Quota**: You create VMs within your quota, from Compute, Isolated machine. The first boot takes a few minutes.
- **Visibility**: Start, stop, restart and remove your own machines from Compute. Removing one means typing its name.
- **Access**: Connect over Remote Desktop with the `.rdp` file from the machine's page. The administrator password is shown once when the machine is created.
- **Resizing**: Stop the machine first. The disk can grow but never shrink.
- **Not included**: No browser file access and no snapshots. Use a container when you need those.
## Life of a virtual machine
Mermaid source of the diagram on this page:
```mermaid
%% Virtual machines: create to delete. Facts from server/routes/vm.js (create, patch, start/stop, rdp, delete), server/tiers.js (kata, Ultra), web/apps/web/components/compute/.
flowchart TB
create["1. Compute, Isolated machine
Ultra plan
name, template, size
within your quota"]
boot["First boot
takes a few minutes"]
run["2. Running"]
rdp["Remote Desktop
download the .rdp file
the administrator password
is shown once at creation"]:::hot
stop["Stopped
change size: the disk can
grow but never shrink"]
del["Delete
you type the name
to confirm"]
create --> boot --> run
run -->|connect| rdp
run -->|Stop| stop
stop -->|Start| run
stop --> del
classDef hot stroke:#8f7ac9
```
---
For graphical work
# Desktop control
An agent that works on your own computer over Tailscale and SSH. What happens there stops one move short of doing the work for you.
## Where it stops
The server prepares the command the agent wants to run and records it, and the exact command is returned to you. It does not execute it for you yet. That applies to your own computer only. Opening SSH or a drive to your own ZeroLog container from Windows runs in one click, see [Drive and SSH in one click](https://zerolog.ch/docs/connect).
**Desktop session** Ends at 02
1. 01
### Composed
The agent works out the command it wants to run.
2. 02
### Shown
The server returns that exact command and records the attempt.
3. 03
### Run
Not done for you. Not yet.
The third compartment is drawn and left empty. That gap is the whole difference between a guided path and a hands-off one.
## What it needs
- **Account access**: The Tailscale permission on your account.
- **Your computer**: Tailscale and an SSH server, registered by its tailnet name, not an IP address.
- **Per session**: Control is granted for one tailnet session and ends with it. The host must also be approved as a tailnet target, and a /goal job cannot approve its own.
- **Actions**: Run a PowerShell command (Windows), run a shell command, fetch a file, take a screenshot. Every attempt is recorded before it runs.
- **A warning**: Anything the agent reads can try to instruct it. With desktop control on, such an instruction would run on your machine. Keep the target list short and switch it off when you are done.
- **Your account**: Scoped to your account, like everything else here.
This page is about reaching your own computer from an agent. The native ZeroLog app for your desktop is a different thing: see [ZeroLog Desktop](https://zerolog.ch/docs/zerolog-desktop).
So **guided** means what it says: the command gets worked out for you, and running it does not.
## Desktop control, step by step
Mermaid source of the diagram on this page:
```mermaid
%% Desktop control: your own computer over Tailscale and SSH. Facts from server/desktop.js, server/routes/desktop.js, server/tailnet.js.
flowchart TB
reg["1. Register your computer
by its tailnet name
(Tailscale and SSH on it)"]
en["2. Enable it for one
tailnet session
the host must also be
approved as a target"]
act["3. Pick an action
PowerShell or shell command
fetch a file, screenshot"]
rec["The attempt is recorded
before anything runs"]
cmd["The server builds the exact
SSH command and returns it"]:::hot
stop["Nothing runs it for you yet"]
end_["When the session ends,
control ends with it"]
reg --> en --> act --> rec --> cmd --> stop
en --> end_
classDef hot stroke:#8f7ac9
```
---
# /goal
Type it and what “done” means, in plain language.
`/goal every test in ~/budget passes`
the checkable end state
*Give it a condition the checker can see in the chat (“pytest reports 0 failed”, “the build prints no errors”). “Make it better” never becomes true.*
## The loop
The checker is a separate, small model. It cannot run anything; it judges only what the chat shows, such as a tool result or an exit code. A plan or a claim of success without a result behind it does not count. Each check is billed like any other model call and shows up in your usage as a goal check; in a free chat the free router does the checking.
The checker sees up to 400 messages since the goal was set, limited to the last 16,000 characters.
It reports impossible only when the model cannot remove the obstacle itself, such as an access right you need to grant. Slow progress does not count.
## What does not interrupt it
the run
- You close the tab
It runs on the server. The next round still starts without you.
- You reload the page
The lines are where you left them.
## What you see
Every round shows up as a line in the chat.
- Stop hook · not met yet
- Hook called model again
- Goal met
## Stop and resume
Stop pauses a goal instead of ending it. Writing the next message does not restart it. The goal bar says paused and shows a **Resume** button. Background goal runs ask for approval like the chat does. See [Stop button](https://zerolog.ch/docs/stop).
With your own key the checker runs on your provider: [helper and checking models](https://zerolog.ch/docs/helper-models).
## Ceilings
### Rounds default 25
Stalls after four checks in a row naming the same missing piece. Otherwise capped at 25 continuations; the operator can change that number.
Need the older, unattended kind instead? `/jobgoal` (`/jg`, `/job`) starts a background job with its own log under [Jobs](https://zerolog.ch/docs/jobs).
### Time per run, by plan
A run lasts at most 4 hours on Mini, 8 on Pro and 12 on Max. Ultra has no limit. The Free plan has no autonomous runs. The clock starts when you set the goal; a new goal starts a new run. At the limit the goal ends with a notice in the chat. History and files stay.
### Credit
With no credit left the goal ends with a notice; top up and set it again. If the checker itself cannot answer, the goal stays set and a message from you continues it.
### Condition
Up to 4000 characters. Chats without an account cannot set a goal.
Plans and prices: [Plans](https://zerolog.ch/docs/plans), [Pricing](https://zerolog.ch/pricing). What else limits spending: [Limits](https://zerolog.ch/docs/limits).
Overview of unattended runs and what limits them: [/goal and /ultracode](https://zerolog.ch/unattended-agents).
### Time, credit and approvals
Standard limits are 4 hours per run on Mini, 8 on Pro and 12 on Max; Ultra has no plan time limit, and the Free plan has no autonomous runs. This is per run, not a monthly hours allowance. See [current plans](https://zerolog.ch/pricing). Paid catalogue requests stop when credit is exhausted. A missing approval, provider error or stopped runtime can prevent progress; an unlimited plan is no guarantee of completion.
When you close the browser, the chat run and stored goal stay on the server. Files, conversation history and a separate terminal process have different lifetimes. See [unattended runs](https://zerolog.ch/unattended-agents), [workspace persistence](https://zerolog.ch/ai-coding-agent-workspace) and [terminal sessions](https://zerolog.ch/docs/terminal).
## How a /goal loop runs and ends
Mermaid source of the diagram on this page:
```mermaid
%% /goal: the stop-hook loop in a chat. Facts from server/zielhaken.js (zielUrteil, zielHaken, pruefeZiel), server/routes/chats.js (watchdog, onDone), web/apps/web/lib/slash.ts.
flowchart TB
set["You type /goal and a condition"]
turn["The model works on a turn
(same run, same tools)"]
check["Evaluator model reads
the chat since
the goal was set"]:::hot
subgraph END["The goal ends"]
direction LR
met["Met"]
imp["Impossible"]
stall["Stalled
same piece
4 checks"]
cap["Capped
25 rounds"]
lim["Time limit
or no credit"]
end
set --> turn
turn --> check
check -->|not met: try again| turn
check --> met
check --> imp
check --> stall
check --> cap
linkStyle 2 stroke:#8f7ac9
```
---
# Stop button
> Stop ends the run, not only the view. One press reaches the tool that is running, the other server slot, the goal and the Ultracode jobs of the chat. The button shows Stopping and…
**Stop ends the run, not only the view.** One press reaches the tool that is running, the other server slot, the goal and the Ultracode jobs of the chat.
- The button shows **Stopping** and stays locked until the server confirms the end. If the request fails, it is sent once more; a second failure shows an error.
- Every tool the assistant can call watches the Stop signal, also tools that wait for a workflow. After a stop, no late result lands in the chat.
- During a release a run can sit on the other server slot. The stop reaches it within a few seconds. A stop that arrives while a message is still being prepared is remembered, and the run starts already stopped.
- Stop **pauses** a [/goal](https://zerolog.ch/docs/goal) instead of ending it. Writing the next message does not restart it. The goal bar says paused and shows a **Resume** button.
- The [Ultracode](https://zerolog.ch/docs/ultracode) jobs of the chat are paused too. Their card shows **Resume**, which continues where the job stopped. A job that fails for a reason a retry cannot change, such as a missing checking model, is not repeated.
We tested this with a replica of the tools, the server slots and the database. We have not yet watched it on a real long build inside a container.
## What one press on Stop reaches
Mermaid source of the diagram on this page:
```mermaid
%% Stop button: what one press reaches. Facts from server/routes/chats.js (/abort), goal pause, job cancel, web stream.ts.
flowchart TB
press["1. You press Stop"]
lock["2. Button shows Stopping
and stays locked"]
subgraph REACH["3. One press reaches"]
direction TB
tool["Running tool ends
no late result lands"]:::hot
slot["Other server slot
gets the stop in seconds"]
goal["/goal is paused
Resume button"]
wf["Ultracode jobs of the chat
pause, Resume on the card"]
tool ~~~ slot ~~~ goal ~~~ wf
end
conf["4. Server confirms the end
button unlocks"]
press --> lock --> REACH --> conf
classDef hot stroke:#8f7ac9
```
---
# /ultracode
`/ultracode` is a switch, not a job. Turn it on, and the model in this chat may split a large task across several agents, with two checkers reviewing each part. It decides when. Short questions still get a short answer.
Type it the same way
```
/ultracode on for this chat
/ultracode on, plus the task
/ultracode off off
```
In the chat, the pill next to the model name opens the Ultracode switch with the reasoning effort slider below it. The pill shows the effort the next turn will use; the steps depend on the engine, and Max carries a note that it uses 1.5 times the usage or more.
The word `ultracode` in a message does not switch anything on for good. It lets the model use the workflow tools for that one turn, and the model decides whether the task fits. If not, it works directly and says why. "No ultracode" or "what is ultracode?" gives no permission. For small mechanical parts see [sub-agents](https://zerolog.ch/docs/subagents). `/swarm ` skips the decision and starts a workflow right away; your account needs the swarm permission for that. In the terminal client, `/ultracode-jetzt ` does the same.
Fan-out
## It fans out
The task is decomposed and the parts run in parallel rather than one after another. Breadth costs wall-clock time once instead of once per part.
One after another
In parallel
Wall-clock
*Both bands carry the same parts and the same work. Only the arrangement differs.*
Refutation
## It reviews adversarially
Each part gets two checkers. Their job is to find what is wrong with the files an agent changed, not with its summary.
The checkers can read files and run commands such as tests, but they cannot write. One looks only for factual or logical errors, the other only for parts that are missing or half-done.
A fixer then repairs the flagged defects once. That fix is not reviewed again: the part is marked as changed after review, and the summary says so. The summary also reports parts where a checker could not answer. A finished workflow does not by itself show that tests passed.
Card
## A card tracks it in the chat
While it runs, a card appears with the plan and one line per agent. The arrow on a line opens its log.
- **The plan**: What the model gave each agent, in its own words
- **Per agent**: Name State A running clock
- **The log**: The task it was given Every tool call What it wrote back
The card streams from the server. Reload, switch devices or close the tab: it picks up where the run is, and the model waits for the result before it answers you.
Status
## A status line shows the work
While a run is active, one fixed line above the message box shows what the agents do: one count (for example 2 of 6 agents active), the stage, the clock and the cost. Open, it lists the agents with their live log. Errors stay visible when the line is closed.
The agents follow the approval setting of your chat. A request shows up in that line with **Allow**, **Deny** and **Allow for this run**. Details: [approvals and status line](https://zerolog.ch/docs/ultracode-approvals).
Files attached to the chat go into the brief of the planner and the agents. A file that is too large for the prompt appears as an outline, its first lines and its path in the container. Images are named in the brief but not described, because the planning models read text only. Only files from your own chat are used.
With your own key you can choose the checking model: [helper and checking models](https://zerolog.ch/docs/helper-models). **Stop** pauses the jobs of the chat and the card shows **Resume**: [Stop button](https://zerolog.ch/docs/stop).
Cost
## It costs more
Several models on one task is several bills. Use it when being wrong is expensive, and [/goal](https://zerolog.ch/docs/goal) the rest of the time.
- **/ultracode**: Audits Migrations Reviews before a release
- **/swarm**: Skip the decision, start right away
- **/goal**: The rest of the time
Ceilings
## The ceilings still apply
It obeys the same ceilings as any other job: your plan's run time, the budget governor and a spend cap if the model sets one.
Every lane
*One bar, drawn across every lane at once. Fanning out buys no exemption from it.*
It needs a paid plan: on the Free plan the switch is not available. The model starts a workflow only while the switch is on for the chat.
Agents use models available to your account. If a model cannot run, follow the message in chat or choose another model from your picker.
Several agents on one task mean several model bills. Credit and plan prices: [Pricing](https://zerolog.ch/pricing); limits: [Limits](https://zerolog.ch/docs/limits).
Overview of unattended runs and what limits them: [/goal and /ultracode](https://zerolog.ch/unattended-agents).
For a continuing finish condition use [/goal](https://zerolog.ch/docs/goal). See [unattended runs](https://zerolog.ch/unattended-agents) and [spend limits](https://zerolog.ch/docs/limits) before delegating long work.
## How an /ultracode workflow runs
Mermaid source of the diagram on this page:
```mermaid
%% /ultracode: one workflow. Facts from server/werkzeuge/ultracodetools.js, server/job-handlers.js (planSubtasks, registerHandler('ultracode')), server/routes/jobs.js.
flowchart TB
on["/ultracode is on for this chat"]
decide["The model decides whether
a workflow pays off"]
plan["Planner splits the task
into 2 to 6 independent parts"]
subgraph PART["Each part, in parallel (8 at once by default)"]
direction TB
impl["Agent does the work
in your container"]
v1["Checker 1
looks for errors"]
v2["Checker 2
looks for gaps"]
fix["Fixer, only if
a checker objected"]:::hot
end
sum["Summary agent reports
to the chat, with the
review status of each part"]
on --> decide --> plan --> PART
impl --> v1
impl --> v2
v1 --> fix
v2 --> fix
PART --> sum
linkStyle 5 stroke:#8f7ac9
linkStyle 6 stroke:#8f7ac9
```
---
# /sub-agents
A sub-agent is one cheap model doing one clear part of a task for the chat model. It exists to save credit: the expensive model plans and checks, the cheap one does the mechanical work. You do not need [Ultracode](https://zerolog.ch/docs/ultracode) for it.
Ask for it in plain words
```
use sub-agents to save money:
read the 3 files in /work/probe and summarise them
```
When
## When the model hands work over
When you want to save credit, or when a part is mechanical: reading several files and summarising them, reshaping data, trying variants. Not for small things it can finish in one step itself.
Setup
## What a sub-agent gets
- **Model**: Cheap by default Your own key: its helper model A dearer model asks you first
- **Tools**: None, or run_command, read_file, write_file in the chat container Chat approvals apply
- **Answer**: 1500 characters by default 6000 at most The chat model checks it
Limits
## Hard limits
At most 4 sub-agents at the same time in a chat and 8 per turn, each with its own spend cap. A sub-agent cannot start more sub-agents or an Ultracode workflow. **Stop** stops them together with the rest of the turn.
Cost
## What it costs
With your own key, nothing for the platform. On the cheap platform tier the cost goes into the normal billing, bounded by the delegation spend cap. Each sub-agent shows as a compact card in the chat with its log.
---
# Ultracode approvals and status line
> Ultracode agents follow the permissions of your chat, and a status line shows their work.
**Ultracode agents follow the permissions of your chat, and a status line shows their work.** An agent runs only what the approval level of the chat would allow.
## Approvals
- The four approval settings of the chat apply: **Ask every time**, **Allowlisted commands only**, **Auto-accept edits** and **Never ask** (the bypass level). One function decides for the chat and for the workers. See [Command approval](https://zerolog.ch/docs/approval).
- A request appears in the status line with the full command and three buttons: **Allow**, **Deny** and **Allow for this run**. The last one remembers the tool for the rest of the run, also after a worker restart.
- Unanswered after five minutes, a request counts as denied. While it waits, no model is called, and Stop still works. Only the owner of the run can answer.
- A short list of protected commands asks on every setting except Never ask: recursive delete outside the work scope, hard resets and git clean, force-push, and anything touching /vault.
- Background `/goal` runs go through the same gate and ask like the chat does.
## The status line
- One fixed line above the message box, with one count, for example 2 of 6 agents active. Closed, it reads like `3 agents · Implementing · 1m 07s · $0.012`.
- Open, it lists the agents with their live log. The log shows each command in full, up to 2,000 characters, with secrets and vault values blacked out ([blackout](https://zerolog.ch/docs/redaction)). It renders Markdown, without HTML and without external images.
- Errors stay visible when the line is closed, for example a missing checking model with the attempt number.
The list of protected commands is a safety net, not a sandbox. The container and the network broker stay in place.
More on the run itself: [/ultracode](https://zerolog.ch/docs/ultracode). Checking models: [helper and checking models](https://zerolog.ch/docs/helper-models).
## What happens to one command of an Ultracode agent
Mermaid source of the diagram on this page:
```mermaid
%% Ultracode approvals: one command of one agent. Facts from server/freigabe-gate.js, routes/jobs.js (freigabe), job_freigaben.
flowchart TB
ask["1. An agent wants to run a command"]
gate["2. The chat's approval level decides
the same gate as in the chat"]
run["Command runs"]
wait["3. Status line shows the full command
no model is called while it waits"]:::hot
deny["Denied
the agent is told"]
ask --> gate
gate -->|"Run without asking, or on your list"| run
gate -->|"Ask every time, not on the list,
or a protected command"| wait
wait -->|"Allow, or Allow for this run"| run
wait -->|"Deny, or no answer in 5 min"| deny
linkStyle 2 stroke:#8f7ac9
classDef hot stroke:#8f7ac9
```
---
One entry per run
# Jobs and logs
Every long-running piece of work appears under **Jobs** with its state and its log.
## What gets an entry
Every long-running piece of work
- A goal or ultracode run
- A research run
- An import or export
- A website build
appears under
Jobs
- its state
- its log
Each entry carries its log, not just a status line. The list shows your latest 100 entries. Finished maintenance runs, such as the background summaries of long chats, are hidden behind a toggle that names how many there are and what they cost.
## From running to what it cost
While it runs
### Follow it live
You can follow a running job live, reading the log while the work is still happening. **Cancel run** stops it after a confirmation; a stopped job is paused, not lost.
the run ends
After it ends
### Read it afterwards
The log stays with the job once the run is finished, so you read it afterwards instead of catching it in the moment. Open an entry to see its trace: model calls, shell commands, tokens, time and cost, and how many requests the broker denied. You can download the trace as JSON. A failed job shows why it failed.
Per job
### What it cost
Cost is attributed per job, so you can see what **a particular run** actually cost.
A job that stopped because credit ran out, an approval is needed or a guardrail tripped waits as paused. **Resume** puts it back in the queue; for a guardrail you can grant extra iterations or budget. Goal and ultracode runs that hit an error retry on their own, up to 10 times.
## Background tasks in a chat
In a chat with a workspace, the model can start a shell command in the background and keep working. It gets a task id at once; the output goes to a file in `/work/.zerolog/hintergrund/`. When the command ends, the chat gets a new turn with the exit code, the duration, the end of the output and the path of the file, even if nobody has the chat open. The model can list its tasks, read the output from any byte offset and stop a task.
- A background command may do exactly what the same command in the foreground may do: same approval, same command list, same network rules, same workspace.
- Limits: 4 running tasks per chat and 8 per account, 30 minutes by default and at most 60, 10 MB of output per task (the command keeps running, the file stops growing), at most 12 wake-up turns per chat and hour.
- A background command can also be a watcher: every new line of output wakes the chat (collected, at most every 30 seconds), and so does the end. A watcher needs a time limit and stops itself after 20 wake-ups.
- The workspace rail of the chat lists running and recent background tasks, wake-ups and recurring instructions. There you read the output, stop a task, and confirm or discard a recurring instruction the model proposed.
- A stopped task does not wake the chat. If the workspace restarts or is offloaded while a task runs, the chat is told that the task ended without a result.
- A wake-up turn is a normal turn and is charged like one. Without credit no turn starts; the result waits, and a note in the chat says so.
## States of a job
Mermaid source of the diagram on this page:
```mermaid
%% Jobs: the life of one entry. Facts from server/jobs.js (runJob catch ladder, cancelJob, versuchsplan), server/routes/jobs.js, web/apps/web/app/(app)/jobs/page.tsx.
flowchart TB
q["Queued"]
r["Running
log grows live"]
d["Done
log and cost stay"]
f["Failed
with the reason"]
p["Paused"]
q --> r
r --> d
r -->|error: retry,
goal and ultracode| q
r -->|out of retries,
or other types| f
r -->|no credit, approval,
guardrail, your Cancel| p
p -->|Resume or Let it
continue| q
linkStyle 5 stroke:#8f7ac9
```
---
# Scheduled tasks
Give an agent a job and a time. A task starts a `/goal` or `/ultracode` run on a schedule. It needs a paid plan.
*For example*
| Job | Time |
| --- | --- |
| A build | Nightly |
| A check | Daily |
| A summary | Weekly |
## Your timezone
Each task has its own **timezone**, and the schedule follows it, not the server's. The form starts with Europe/Zurich; change it to yours. Daylight saving is handled. Pick a preset (daily at 03:00, daily at 08:00, every hour, Mondays at 03:00) or type a five-field cron expression, and the form shows the next three run times before you save.
## The cap, if you set one
A schedule has **no spend cap by default**. You can set a budget on the task; a run then stops at it. Without one, a run uses credit until it ends or your credit reaches zero. See [Limits](https://zerolog.ch/docs/limits).
## The log
Every run leaves a **log you can read in the morning**: it is a job under [Jobs](https://zerolog.ch/docs/jobs), with its trace and cost. If the worker was down while a run came due, the task fires once when it is back, not once for every missed time. A task stops by itself at its maximum number of runs.
## From a schedule to a run
Mermaid source of the diagram on this page:
```mermaid
%% Scheduled tasks: from schedule to a run. Facts from server/cowork.js (nextRun, runDue, retire, startScheduler), server/routes/tasks.js, web/apps/web/app/(app)/tasks/page.tsx.
flowchart TB
sched["You create a task:
name, type (goal or ultracode),
schedule, timezone
optional: budget per run,
maximum runs"]
tick["The worker checks every
30 seconds which tasks are due"]
ok{"Still allowed?"}
stop["Task is switched off:
maximum runs reached,
or the permission is gone"]
job["A job is queued
with your budget as spend cap"]:::hot
jobs["The run appears under Jobs
with its log and cost"]
next["Next run time is computed
in your timezone"]
sched --> tick --> ok
ok -->|no| stop
ok -->|yes| job --> jobs
job --> next
next -.-> tick
linkStyle 3 stroke:#8f7ac9
```
---
# Invisible automation
Reading needs no permission. Anything that leaves the system needs a standing grant from you and one first confirmation. Money is never automated. What is not allowed waits in an approval queue.
## What a grant is
You grant one action type at a time with a switch in Settings, Automation: send mail, reply to mail, label or archive mail, create or change a calendar event, create or edit a document, call an outside webhook. Grants are **explicit and revocable**, so nothing gains permission **silently**. Revoking one also clears its first confirmation, so granting it again asks again. The feature needs a paid plan.
## Why the digest exists
“It ran while I was away” never means “I do not know what happened.”
Findings are collected and delivered once per period, as a plain list with the actions still waiting for you. Nothing pings per event, and a day with nothing to report sends nothing.
## How long it can run
Your [plan](https://zerolog.ch/docs/plans) sets the running time for unattended background jobs. Follow a job live, or read its log afterwards, under [Jobs](https://zerolog.ch/docs/jobs).
For a `/goal` in chat, the automatic limits are rounds and repeated lack of progress.
## Wake-ups and recurring instructions from the chat
The model can set a one-time wake-up for the chat it is in: after the time it chose, the chat gets a new turn with the model's own note. A wake-up lies between one minute and seven days ahead, a chat holds at most five.
The model can also propose a recurring instruction for the same chat. It does not run until you confirm it in the background tasks panel. Every proposal names how often it runs (at most every 15 minutes), how many runs it may make (at most 100), a spending cap for the chat (at most CHF 50) and an end date (at most 30 days). When the chat reaches the cap, the schedule ends and a note says so. You can delete any wake-up or schedule at any time.
## How an automation action is gated
Mermaid source of the diagram on this page:
```mermaid
%% Automation gate: what happens to an action an automation wants to take. Facts from server/automation.js (ACTIONS, check, propose, approve, digest), server/routes/automation.js.
flowchart TB
act["An automation wants
to take an action"]
kind{"What kind?"}
read["Read only
mail, calendar, documents,
a web page"]
write["Leaves the system
send or reply, label mail,
calendar, documents, webhook"]
money["Money
transfer, purchase, refund"]
run["Runs, no grant needed"]
grant{"Standing grant,
confirmed once?"}
allowed["Runs"]
queue["Waits in the
approval queue"]:::hot
never["Refused, always"]
you["You approve or reject.
Approving also records the
first-time confirmation"]
act --> kind
kind --> read --> run
kind --> write --> grant
kind --> money --> never
grant -->|yes| allowed
grant -->|no| queue --> you
linkStyle 8 stroke:#8f7ac9
linkStyle 9 stroke:#8f7ac9
```
---
# Website Builder
Describe a website in one sentence. Get a built, checked project with a preview and export bundles.
The website builder turns a short description into a built, checked and exportable web project. `/site ` starts it as an unattended background run: it appears under [Jobs](https://zerolog.ch/docs/jobs) with its log and cost, and you can close the chat while it works. The template previews appear in the chat. `/design` is the other route, where the AI designs with a live preview next to the chat.
## What a run does
| Step | What it does | Your part |
| --- | --- | --- |
| 1. Brief | Turns the description into a brief. It never asks questions back; every assumption and every gap is written down and listed in the final report. | None. |
| 2. Research | Fills open facts, but only when a quote from a fetched source confirms them. Otherwise a visible placeholder stays. | None. |
| 3. Template previews | Picks a style from 181 curated ones and shows five template previews in the chat. The best-scored one is used at once. | Optional: choose another. The build restarts at the content step. |
| 4. Annotation | Draw an arrow, circle, frame, freehand stroke, text note or point on the preview to ask for a precise change. | Optional. |
| 5. Build | Writes content, scaffold, images, frontend, backend and wiring in your container. | None. |
| 6. Checks | A build gate (install, typecheck, lint, build, boot, smoke test); a visual critic that takes screenshots at 360, 768 and 1440 px and repairs defects; a security gate (code rules, dependency audit, response headers); an acceptance check that renders the served page and repairs what does not work. A site that does not build is not exported. | None. |
| 7. Finish | Suggests domain names and checks them, builds the export bundles and writes a report. | Download a bundle, or deploy it. |
Visual markup notes stick to HTML element selectors rather than screen coordinates. If a revision moves a button or re-flows a grid, your note stays attached to the element you clicked. If no selector can be found, the note says so.
## Options
Flags go after the description: `--profile ssr` (a server-rendered app instead of static pages plus API), `--budget `, `--lang de,en` (up to four languages), `--no-domain`, `--export docker,static`, `--model ` and `--firma `. An unknown flag stays part of the description. Starting a run returns a cost estimate. For several sites in one go, see `/ultrasite`.
## Export bundles
Seven targets exist: `docker`, `static`, `webspace` (classic hosting), `zerolog-native` (the site stays where it was built), `vercel`, `netlify` and `cf-pages` (Cloudflare Pages). Without `--export` you get docker, static, webspace and zerolog-native; with `--profile ssr` only docker and zerolog-native. Every bundle comes with a README, an `.env.example` and a `VERIFY.md`. Vercel, Netlify and Cloudflare Pages host the frontend only; the backend runs from the docker bundle.
## Distinct design directions
When Design Engine v2 is available for your build, it compares each direction with your earlier builds and common design patterns. If a direction is too similar, the builder revises it, at most twice, before it creates the full site.
Completed v2 builds add design and review signals to a rolling memory of your last 50 builds. Deleting a project also deletes its saved design memory.
## Build budget and pauses
If you set a build budget, the builder pauses before starting another model request once recorded spending reaches it. Completed responses stay saved, so you can continue from the saved progress.
To continue after the budget is exhausted, open the paused build in Jobs, enter a positive amount in **Extra budget (CHF)**, then choose **Resume**. ZeroLog records the added budget. If the amount is missing or too small to raise the limit, the build stays paused.
Requests already in progress can finish above the budget. Their usage still counts toward the total, so the budget does not guarantee an exact maximum charge.
## What to check yourself
For Swiss businesses the builder adds imprint and privacy pages, and notes that the legal texts must be checked by the operator. Placeholders that remain visible are facts it could not confirm.
## How the website builder works
Mermaid source of the diagram on this page:
```mermaid
%% Website builder: the /site pipeline. Facts from server/site-builder/plan.js (NODE_ORDER), pipeline.js, nodes/*.js, export/index.js.
flowchart TB
brief["Brief
one sentence in, assumptions written down"]
research["Research
facts only with a quoted source"]
sty["Style and 5 template previews
shown in the chat"]
pick["The best-scored template is used at once.
If you pick another later, the build restarts at content"]:::hot
build["Build in your container
content, scaffold, images,
frontend, backend, wiring"]
gates["Four checks in order
1 Build gate: install, lint, build, boot
2 Visual critic: 360, 768, 1440 px, with repairs
3 Security gate: code rules, audit, headers
4 Acceptance: the served page is rendered and repaired"]
out["Domain ideas, export bundles, report"]
brief --> research --> sty --> pick --> build --> gates --> out
linkStyle 3 stroke:#8f7ac9
```
---
# Browser extension
Use ZeroLog chats in your browser side panel. Ask for help with pages you can access, then follow the activity in the panel.
## Install and pair
Install the Chrome or Firefox build from [the extension page](https://zerolog.ch/extension). In the extension, choose **Connect with ZeroLog** and approve in your signed-in account. The one-time code from the extension page still works as well.
## Connecting
**Connect with ZeroLog** pairs in one click. The extension keeps a secret verifier and zerolog.ch only sees its challenge. You approve in your signed-in account, and the extension finishes the pairing once. Code pairing still works. You can set a default browser per account and test the connection.
## Use the side panel
Open a ZeroLog chat in the panel and choose the browser tab you want help with. The model can read page content and carry out browser actions. The panel shows its activity, and you can review and approve actions when prompted.
When filling forms, ZeroLog will not enter passwords, payment-card details or one-time codes. Enter those yourself.
Page content used to answer your request follows the same retention settings as your chats. Read [Zero data retention](https://zerolog.ch/docs/zdr).
## The panel in version 0.3
The panel has four tabs: Chats, Page, Browser and Connection. Answers show as Markdown and you can search your chats. The Browser tab lists the current task from the real tool steps, and a banner offers to take over. After installation a setup page opens, and text you select on a page reaches the panel through the context menu.
## The agent's own tab
Since version 0.3.1, if no agent tab is open, `browser_navigate` opens its own tab in a tab group called ZeroLog and keeps using it. The extension asks for the tabGroups permission for this.
The safety rules stay: no sensitive fields, **http and https** pages only, the site rule below, **45 seconds** per action, and a visible stop button.
## What needs your approval
The permission tier, set in the Connection tab, decides which actions show an approval card. **Smart** is the default: reading the page and navigating run on their own; clicks, typing, key presses and tab changes ask first. **Manual** asks for every action. **Always** never asks.
The tier never lifts the hard rules: passwords, card details, one-time codes and CAPTCHAs stay off limits, and only http and https pages can be opened.
The site rule is a setting in ZeroLog. By default every site is allowed unless you blocked it. In strict mode a site works only after you approve it, and that approval stays until you remove it. A block always wins in either mode.
## Choose a connected browser
You can pair Chrome and Firefox and choose which connected browser to use from the panel. If one browser closes during a task, the extension can continue in another connected browser.
## Updating, and Firefox
The Chrome build does not update itself. Download the new package from the extension page, replace the folder and reload the extension on chrome://extensions.
For Firefox, install the extension from [Firefox Add-ons](https://addons.mozilla.org/firefox/addon/zerolog-browser-extension/). It needs **Firefox 140 or newer** (Chrome 116 or newer for the Chrome build), stays installed after a restart and receives updates through Firefox. The extension page also offers the file directly: a Mozilla-signed copy when it matches the current extension version, otherwise the unsigned source download as a developer preview.
## Send files in a chat
Use the **+ file** button or drop a file into the panel. Images are sent as images when the selected model supports them. Text files are added as text; other files are attached by name.
Up to **8 files** per message, **20 MB** per file and **24 MB** per message. Files and replies stay in the same chat you can open in the web app.
## Your chats
Continue an existing ZeroLog conversation or start a new one in the side panel. Files downloaded during a task appear in the panel's download list.
## What happens to one browser action
Mermaid source of the diagram on this page:
```mermaid
%% Browser extension: one browser action. Facts from server/extrelay.js, extension/tools.js, extension/sw.js, extension/lib/guards.js, page-input.js, server/bridges.js.
flowchart TB
chat["You ask in the side panel chat"]
model["The model asks for a browser tool
navigate, click, type, read, screenshot ..."]
relay["ZeroLog relays it to the
extension of the connected browser
(45 s per action)"]
subgraph EXT["Four checks inside the extension"]
direction TB
c1["Only http and https pages"]
c2["Site rule: open by default,
or strict, where each site needs your OK"]
c3["Never into password, card
or one-time-code fields"]
c4["Permission tier decides
if an approval card is needed"]:::hot
end
run["The action runs on the tab
(navigation opens a tab in the
ZeroLog tab group)"]
back["Result and screenshot
go back to the model"]
c1 ~~~ c3
c2 ~~~ c4
chat --> model --> relay --> EXT --> run --> back
```
---
# Trust and security
Who runs ZeroLog, how a workspace is isolated, where the service is hosted, and which outside parties see your data.
## Operator and contact
ZeroLog is run by Gian-Luca Luongo, a sole proprietor (natural person) based in Wettingen, Switzerland. Contact for any question, including security reports, is [info@zerolog.ch](mailto:info@zerolog.ch). Full legal details are in the [Impressum](https://zerolog.ch/legal#impressum). Vulnerability reports follow [security.txt](https://zerolog.ch/.well-known/security.txt) (RFC 9116).
## Architecture, from your workspace's side
- Each workspace is its own container, not shared with any other user: a gVisor container, or on Ultra optionally a Kata micro-VM with its own kernel — see [Isolation](https://zerolog.ch/docs/isolation).
- Third-party API keys you store are not kept in the main application. A separate key service on its own Swiss VM holds them encrypted; agents only see placeholders — see [Key proxy](https://zerolog.ch/docs/keyproxy).
- Outbound traffic from a workspace runs through a filtered network path — see [Outbound network](https://zerolog.ch/docs/egress).
- A command the AI wants to run waits for your approval first — see [Command approval](https://zerolog.ch/docs/approval).
- Chats, files and workspace disks are stored on the server in Seon (Aargau), Switzerland; pausing a workspace keeps its disk until you delete it yourself.
## Where prompts go
Prompts and the files or images you attach are sent to Featherless, based in the USA, to generate a reply: this is the model inference behind every customer account and is the core of the service. Optional connectors only receive data if you switch them on.
## Who receives data
| Recipient | Location | Purpose |
| --- | --- | --- |
| Featherless and Global DPA Annex III subprocessors | USA, Singapore and other locations without a region guarantee | Model inference; provider-internal fallback possible |
| Stripe | Ireland and USA | Payment processing |
| Cloudflare | USA; edge location nearest you | Reaching the site, DDoS protection |
| IONOS SE | Germany | Outbound network gateway for containers; encrypted offsite backup copies, without the decryption key |
| SMTP2GO | New Zealand company; sending locations depend on provider routing | System email; incoming operator mailbox is hosted in Switzerland |
| Google | USA | Optional Gmail, Drive and Calendar connectors, only if enabled; sign-in uses the operator’s Keycloak |
| Infomaniak | Switzerland | Image generation, only when you request an image |
| MoonPay Commerce / Helio, Bybit | See the Privacy Policy | Manual credit top-ups; the payment settles to the operator’s custodial Bybit USDC wallet |
| Anthropic | USA | Automated, redacted error reports |
Featherless confirmed on 29 September 2026 that the Global DPA automatically applies to ZeroLog’s Developer API account. Its EU SCC Module 2 covers EEA/UK transfers; no Swiss Addendum has been agreed. The operator relies on Art. 17(1)(b) FADP for Swiss transfers, with unresolved legal uncertainty for regular transfers. The Developer plan remains in use, without a processing-location guarantee or per-account routing control. The [Privacy Policy](https://zerolog.ch/legal#privacy) explains the recipients, provider commitment and remaining risks.
## Limits, status and changes
- Compute and credit limits are set per plan — see [Plans](https://zerolog.ch/docs/plans) and [Spend limits](https://zerolog.ch/docs/limits).
- Live model-provider status is at [/status](https://zerolog.ch/status); platform uptime history is on the [operator's status page](https://zerologch.statuspage.io/).
- Changes to the service, including security-relevant ones, are announced in the [Newsroom](https://zerolog.ch/news).
## Where your data goes
Mermaid source of the diagram on this page:
```mermaid
%% Trust: where your data goes. Facts from the Privacy Policy recipient list, docs/isolation, docs/keyproxy.
flowchart TB
you["You
browser, CLI or app"]
cf["Cloudflare edge
(USA) reaches the site"]
app["App server, Switzerland
Seon (Aargau)
chats, files, workspace disks
sign-in via own Keycloak"]
ws["Your workspace, Switzerland
own container, approval
before each command"]
ks["Key service, Switzerland
separate VM, keys encrypted"]
fl["Featherless and subprocessors
USA and others
prompts, files, images"]:::hot
pay["Stripe, MoonPay
payments"]
infra["IONOS (DE), SMTP2GO
egress gateway, backups,
system email"]
opt["Anthropic, Google
redacted error reports;
Google only if connected"]
you --> cf --> app
app -->|starts, stops| ws
ws -->|placeholders| ks
app -->|model inference| fl
app --> pay
app --> infra
app --> opt
fl ~~~ pay
pay ~~~ infra
infra ~~~ opt
classDef hot stroke:#8f7ac9
```
---
# Zero-downtime deploys
A release starts the new code next to the old code and then moves traffic over. You are not sent through a restart: a reply that is already streaming, or a file upload that is already under way, finishes on the old copy.
## How a release goes
1. **Test in isolation.** The new code is tested in an isolated copy first: lint and the test suite. Only green code moves on.
2. **Back up.** A backup is taken before anything on the live system changes.
3. **Check the database changes.** Old and new copy share one database for a few moments, so a release that drops or renames a table or column, or adds a required column without a default, is refused.
4. **Start the idle slot.** It starts with the new code. If it does not pass its health check, the active slot keeps serving and the release stops there.
5. **Move traffic.** The new slot is activated and the front-door proxy is told to send traffic to it. If the proxy's own code changed, a second proxy instance starts, is health-checked, and the port redirect is moved to it before the old instance stops.
6. **Drain the old slot.** Follow-up requests of chats and uploads that began on it are kept on it until they finish; it is stopped after at most ten minutes.
7. **Restart the worker.** The background worker restarts and finishes its running jobs first, for at most five minutes.
8. **Smoke test.** A smoke test runs against the slot that is now active.
## What it does not do
A reply that is still running when the old copy's ten minutes are up is cut off. A release that removes or renames database columns is refused in the normal path and needs a planned two-step release instead. None of this is something you configure.
## Going back
The operator can return to the single-process setup. The old process starts and is health-checked while the proxy keeps serving; only then does traffic move to it, and the proxies and slots are removed last.
Related: [Isolation](https://zerolog.ch/docs/isolation), [Trust and security](https://zerolog.ch/docs/trust), [Status](https://zerolog.ch/status).
## How a deploy switches ZeroLog to new code without downtime
Mermaid source of the diagram on this page:
```mermaid
%% Zero-downtime deploy. Facts from scripts/deploy.sh, scripts/zdt-proxy.mjs, scripts/zdt-nft.sh, scripts/zdt-worker-cutover.sh, scripts/zdt-rueckbau.sh.
flowchart TB
visitors["Visitors, tunnel and
worker machines
(all use port 8080)"]
nft["nftables table zl_zdt
redirects port 8080
to the active proxy"]
subgraph PX["Front-door proxy (Node.js)"]
direction LR
pa["zerolog-proxy@a
port 8084"]
pb["zerolog-proxy@b
port 8085"]
end
subgraph SL["App slots, one active"]
direction LR
blue["zerolog@blue
port 8082"]
green["zerolog@green
port 8083"]
end
deploy["deploy.sh
tests, backup, switch"]
worker["zerolog-worker
background jobs"]
visitors --> nft
nft -->|one instance at a time| PX
PX -->|active slot only| SL
deploy -->|start idle slot,
activate| SL
deploy -->|reload proxy| PX
deploy -->|restart,
with drain| worker
linkStyle 3 stroke:#8f7ac9
```
---
# Deep research
Background job · minutes, not seconds
A longer, structured version of [research mode](https://zerolog.ch/docs/research) for questions where you want a report rather than an answer.
## How a run goes
1. \[1\]
### Plan
A strong model turns your question into the pieces the run works from.
- Sub-questions.
- A set of facets to investigate.
- The criteria that would mean the question has actually been answered.
2. \[2\]
### Gather, in parallel
One agent per facet
Facet
Facet
Facet
and so on
Each agent finds sources, fetches them, and pulls out specific claims with the quote each one rests on.
3. \[3\]
### Decide whether to go again
If the agents report open gaps, the planner proposes one to three more facets and the run gathers again. It stops when one of these is true:
- No gaps are left open
- It has run for 90 minutes (the default time limit)
- It has done four rounds
4. \[4\]
### Write the report and check the citations resolve
The report is written from the numbered sources. Every `[n]` in it must point at a source that was actually fetched, and every key finding must carry at least one.
- \[1\] a source the run fetched **holds**
- \[2\] a source the run fetched **holds**
- \[3\] nothing that was fetched **fails**
- \[ \] a key finding carrying no citation at all **fails**
Either failure sends the report back. It is regenerated once; if a problem remains, the report ends with a warning line saying how many citation issues could not be resolved.
The dashed return is the round that goes again: it starts back at gathering, with the extra facets.
## What comes back
The report
Citations you can follow.
Alongside it
A record of what was dropped along the way.
It appears under [Jobs](https://zerolog.ch/docs/jobs) like other long-running work, and survives the browser being closed.
## What it costs
**Several models** *×* **many fetches** *×* **multiple rounds**
This is one of the more expensive things you can run. You cannot set a spend cap for a run: it is billed from your credit like any model use and stops at zero credit, after four rounds or after the time limit. Check your balance before an open-ended question. Deep research needs Pro, Max or Ultra.
## Not the same guarantee as research mode
### [Research mode](https://zerolog.ch/docs/research)
- **Checks**: That a quote appears verbatim in the page it cites.
- **Strength**: The stronger check.
- **Reach for it when**: A single claim has to be bulletproof.
### Deep research this page
- **Checks**: That every citation resolves to a source it really fetched, and that nothing is left uncited.
- **Strength**: A weaker guarantee, on far more material.
- **Reach for it when**: You need breadth and will read the sources yourself.
## How a deep research run goes
Mermaid source of the diagram on this page:
```mermaid
%% Deep research: one run. Facts from server/deepresearch.js, research-job.js, plan-gates.js.
flowchart TB
q["Your question
(Pro, Max or Ultra)"]
p["1. Plan
sub-questions, facets,
success criteria"]
g["2. Gather, one agent per facet
fetch up to 6 pages each,
extract claims with quotes"]
gap{"3. Gaps left?
under 4 rounds,
under 90 minutes"}
s["4. Write the report
from numbered sources"]
v{"Every [n] resolves,
every finding cited?"}
fix["Regenerate once"]
out["Cited report in your chat
and under Jobs
(warning line if issues remain)"]:::hot
q --> p --> g --> gap
gap -->|yes, plan more facets| g
gap -->|no| s
s --> v
v -->|yes| out
v -->|no| fix --> out
classDef hot stroke:#8f7ac9
```
---
# Legal research
Ask a legal question and ZeroLog searches supported primary legal sources. Answers include citations you can open and review.
Access depends on your account permissions.
## Source coverage
The sources are a fixed, shared collection of primary legal texts, not the live web. It holds Swiss, French, Italian, EU, US and Canadian material, in different depth. Coverage varies by jurisdiction and topic.
Before an answer is written, the retrieved passages have to match what your question is actually about. If they do not, ZeroLog says the available sources do not cover your question instead of presenting an unsupported answer.
## Review the answer
Open the citations and read the source text for yourself. Use the answer as a starting point for research, not as a decision about your situation.
Before you act
This is legal research, not legal advice. The answer does not know your full facts and has not been reviewed by a lawyer. ZeroLog adds this notice to every answer itself. Speak with a qualified lawyer before acting on it.
## How a legal question is answered
Mermaid source of the diagram on this page:
```mermaid
%% Legal research: one question. Facts from server/zerolog.js, legalcoverage.js, routes/legal.js.
flowchart TB
q["Your legal question
(needs the permission)"]
s["Search the shared corpus
CH, FR, IT, EU, US, CA"]
c{"Passages really about
your question?"}
a["Answer with citations
you can open and review"]:::hot
n["Notice added by ZeroLog:
not legal advice"]
no["Says the sources do not
cover your question"]
q --> s --> c
c -->|yes| a --> n
c -->|no| no
classDef hot stroke:#8f7ac9
```
---
# Security scan
A security scan checks a website you are authorised to test and reports findings you can review.
1. Checks the authorised website
2. Reviews potential issues
3. Records scan findings
Goes again
Leaves the loop
Validated findings include the reproduction from the scan so you can assess what was detected.
## Proving the domain is yours
ZeroLog checks that you control the domain before allowing a scan.
Security scan *›* Prove domain ownership
Follow the on-screen instructions to confirm control of your website. You publish a DNS record that ZeroLog names: a TXT record at `_zerolog-verify.` plus your domain, or a CNAME at the same name. An A record alone is not accepted. The record is checked again later.
Authorisation matters
If control can no longer be confirmed, scans for that domain cannot continue. Only scan websites you own or have permission to test.
## Running one
1. **Add and verify the domain.**
2. **Pay for the scan.** It is a service you buy each time, not an unlock you buy once.CHF 20 per scan · card, TWINT where enabled
3. **Start it.** It runs as a background job, so you can close the browser.
4. **Read the findings.** Each one carries the proof-of-concept that established it.
Scans run from your own workspace, so you need a plan with a container (Mini or higher). Besides a verified domain, a scan can target your own workspace or a site made with the site builder. Scanning a guest VM is not available yet. You choose how deep it goes: quick, standard or deep, and deep is the default.
Model usage during the scan is billed from your ordinary credit on top of the scan fee, priced per request like everything else.
If the target workspace stops during a scan, that scan cannot finish. ZeroLog reports the scan error and stays available; start the workspace and run the scan again.
## Scope
The scan is limited to the authorised website. ZeroLog applies safeguards intended to keep a scan within that scope: requests to a domain go out on a dedicated route that reaches only verified domains and their subdomains.
After a finished scan of your own workspace or a built site, a separate fix job can patch the findings marked fixable in your own code and scan again. For a domain scan the fixes are guidance only, because there is no code of yours to change.
**Scan only what you own.** Only scan websites you own or have permission to test.
## From domain to findings
Mermaid source of the diagram on this page:
```mermaid
%% Security scan: from domain to findings. Facts from server/pentest-domain.js, pentest-pay.js, pentest.js, routes/pentest.js.
flowchart TB
d["Add your domain"]
dns["Publish the DNS record ZeroLog names
TXT or CNAME at _zerolog-verify"]
v{"Control confirmed?"}
stop["No scan; checked again later,
scans stop if the record goes"]
pay["Pay CHF 20 for this scan"]
job["Background job in your workspace
plan, act, observe, repeat
only verified hosts reachable"]
f["Finding counts only with
a working proof-of-concept"]:::hot
r["You read the findings"]
fix["Optional fix job
own workspace or site only"]
d --> dns --> v
v -->|no| stop
v -->|yes| pay --> job --> f --> r
r --> fix
classDef hot stroke:#8f7ac9
```
---
# Live web search and sources
Search the web from a conversation and review citations alongside the answer.
Grounded web search retrieves current pages and checks that cited quotations match their source text.
- **Direct web retrieval:** Searches ZeroLog's own index and fetches pages you or the model name directly. Results depend on what could be retrieved.
- **Guarded fetching:** A fetch follows your account's outbound-network rules. Every redirect is checked again (at most three), internal and private addresses are refused, and a page over 2 MiB or a fetch over 20 seconds fails. A failed page is recorded with its reason instead of being hidden.
- **Your snapshots stay yours:** A fetched page is kept as a snapshot for your chat. Pages fetched for you are not added to the shared search index.
- **Supported citations:** A citation is marked supported only if its quote (at least 12 characters) appears in the page text we retrieved (ignoring case and whitespace differences). A quote that does not appear is flagged as unsupported.
Search results and citations depend on the sources available at the time of the request. Review cited pages before relying on a result.
## From question to a checked citation
Mermaid source of the diagram on this page:
```mermaid
%% Grounded search: from question to checked citation. Facts from server/research.js (fetchSource, verifyClaim), searchrouter.js.
flowchart TB
q["Your question"]
f["Find pages
ZeroLog's own index, or
URLs you or the model name"]
g["Fetch under your account's
outbound rules, each redirect
checked, 2 MiB, 20 s"]
snap["Snapshot kept for your chat
not added to the shared index"]
m["Model answers with exact quotes
(at most 25 words)"]
c{"Quote found in the
retrieved page text?"}
ok["Supported"]:::hot
bad["Unsupported, flagged
with the reason"]
q --> f --> g --> snap --> m --> c
c -->|yes| ok
c -->|no| bad
classDef hot stroke:#8f7ac9
```
---
# Cost per answer
Every reply carries its own cost line, right under the answer: input tokens, cached tokens, output tokens, and the price. The figures update while the answer streams.
The line under a reply
Answer
**Input tokens**
**Cached tokens**
**Output tokens**
**Price**
**Cached tokens**
Cached input is cheaper than fresh input. Long threads benefit from that automatically.
## The full ledger
The line under a reply covers that one answer. Usage holds the full ledger, per chat, per model and per day.
The same entries, grouped three ways
**Per chat**
**Per model**
**Per day**
---
# Credit and top-ups
Credit is prepaid. You name an amount in CHF and pay through the MoonPay checkout, which settles to the operator’s USDC wallet, currently custodial at Bybit. A card works there, through MoonPay's onramp.
A top-up is never charged automatically. The [paid plans](https://zerolog.ch/docs/plans) add an allowance each month, and credit packs are monthly subscriptions (Stripe, by card, renewing until cancelled) that can be taken on any plan.
Credit is only added when the payment provider's signed confirmation arrives, never when you press pay. One top-up can be at most CHF 5000. Credit from a plan or a pack expires three months after it was added and is spent first; credit from a top-up does not expire.
- **Solid line**: a figure that stands still
- **Dashed line**: moves with the exchange rate
## The smallest top-up
*Fig. 1 · the floor, and the clearance kept above it. Not to scale.*
CHF quote*current*
The minimum in CHF moves with the exchange rate. Checkout calculates the current amount and shows it before you pay.
**clearance**
The payment quote accounts for the minimum and exchange-rate movement.
USD 20
MoonPay's onramp refuses anything under this.
refused
The quote tells you the exact figure before you pay.
## Where the percentages sit
*Fig. 2 · both brackets span everything. Whatever the percentage does, the whole figure is in front of you. Not to scale.*
Topping up
**the amount you name** **3% currency buffer**
the quote, shown before you pay
Topping up adds no service fee. The buffer covers the rate moving between the quote and the money arriving.
Spending it on tokens
**the provider's price** **2% service margin**
the price you see
The margin is already inside every price shown to you: in the picker, in the catalogue, and in the cost strip under each reply.
A billing error caused some non-streaming coding-agent calls to include the service margin twice. This has been corrected.
## Paying from your own wallet
Paying from a wallet you already hold is faster, and it skips the repeated ID check. If you do not have one yet, [Wallets, and how to get USDC](https://zerolog.ch/docs/wallet) covers what you need.
Credit is not redeemable for cash while the account stays open. On account closure, unused purchased credit is refundable under [Terms §12.3](https://zerolog.ch/legal#tos-12-3).
Plan prices and credit packs are listed on [Pricing](https://zerolog.ch/pricing). Changes to prices and billing are announced in the [Newsroom](https://zerolog.ch/news).
## From top-up to spending
Mermaid source of the diagram on this page:
```mermaid
%% Credit: top-up to spend. Facts from server/routes/billing.js, currency.js, moonpay.js, routes/moonpay.js, pricing.js.
flowchart TB
a["You name an amount in CHF
at most CHF 5000"]
q["Quote shown before you pay
amount plus 3% currency buffer"]
m["MoonPay checkout
card, or USDC on Solana
from your wallet"]
w["Settles to the operator's
custodial Bybit USDC wallet"]
h["Signed confirmation
from the provider arrives"]
c["Credit added to your balance"]:::hot
s["Each model request
is paid from it, at provider
price plus 2% margin"]
z["At zero, paid models stop"]
a --> q --> m --> w --> h --> c --> s --> z
classDef hot stroke:#8f7ac9
```
---
# Wallets, and how to get USDC
The payment page supports USDC on Solana. Available wallet and card routes, fees and identity checks are shown by the payment provider before you pay. Confirmation time depends on the network and the provider; an immediate credit is not guaranteed.
## What a wallet actually is
In a self-custody wallet, you control the signing keys. In a custodial account, a provider controls them for you. The distinction concerns your wallet: ZeroLog's merchant settlement uses a custodial Bybit account. See [Payments and privacy](https://zerolog.ch/legal#pp-3-3).
One key pair, two faces of the same fact
The public half
The address. It is what an exchange withdraws to.
The private half
The rest of the pair, sitting on your device in the extension or the app.
Control your signing keys.
Keep your recovery method safe.
The same sentence, read from two sides
Write the recovery phrase down on paper and keep it away from the computer.
Self-custody does not remove token-issuer restrictions. Circle can block USDC addresses under its [USDC terms](https://www.circle.com/legal/usdc-terms). ZeroLog cannot recover a lost private key or reverse a blockchain transaction.
## USDC on Solana, not the other USDC
The network matters as much as the coin. A wallet that speaks Solana is what you want: Phantom and Solflare are the common ones.
- **USDC** Solana *This is the one you need*
- **USDC** Ethereum
- **USDC** Base
- **USDC** Polygon
USDC exists on Ethereum, Base, Polygon and others too, and those are different tokens as far as the payment is concerned.
## Where to buy it
Before using an exchange or onramp, check its eligibility rules, fees, withdrawal limits and privacy terms. The payment route also requires both of these:
- It sells USDC.
- It lets you withdraw to a Solana address.
If you already use a provider, check that it supports the required withdrawal. Choose **Solana** as the network and verify the address on the payment page.
## Identity checks depend on the provider
A provider may request identity documents or further checks based on its rules, your country and the transaction. ZeroLog does not promise that checks happen only once, on every payment, or never.
From a wallet
Your exchange and the payment service may apply their own checks.
MoonPay card route
The provider tells you which identity checks apply before completing payment.
## A transfer cannot be undone
Sending to a wrong address loses the money, and no support desk can reverse it.
Send a small amount first, the first time you use a new address.
Network, exchange and payment-provider fees can differ. Review the current quote and any minimum payment before approving a transfer.
## Topping up, once the USDC is there
1. 01
Pick the amount.
2. 02
Open the payment page.
3. 03
Approve the transfer in your wallet.
The credit appears when the payment confirms.
## Getting USDC to the payment page
Mermaid source of the diagram on this page:
```mermaid
%% Wallet: paying a top-up from your own wallet. Facts from the docs page, server/moonpay.js (USDC on Solana).
flowchart TB
b["Buy USDC at an exchange
or onramp (your choice)"]
n{"Can it withdraw USDC
to a Solana address?"}
other["Pick another provider;
USDC on other chains
does not work here"]
w["Withdraw to your wallet
network: Solana"]
p["Pick the amount,
open the payment page"]
ok["Approve the transfer
in your wallet"]
conf["Provider confirms;
credit appears"]:::hot
b --> n
n -->|no| other
n -->|yes| w --> p --> ok --> conf
classDef hot stroke:#8f7ac9
```
---
# Spend limits
What stops spending on ZeroLog, and what does not. Your credit is the hard stop.
the line you set
## At the limit, work stops
the limit
- Your credit
- Delegated work, per chat turn nothing runs here
- /goal run time, per run
*Three limits that are built. At any of them, work **stops**: it does not continue and bill you later. Paid models stop at zero credit. Delegated sub-agents draw from a default cap of CHF 0.50 per chat turn, with at most four delegated calls. A /goal run ends at the time limit of your [plan](https://zerolog.ch/docs/plans). A single chat turn that is still working is ended after six hours; the work so far is kept and "continue" carries on.*
what is not there
## No default spend cap per run, per chat or per agent
No plan sets a spend cap per run by default, and there is no spend limit per chat. A long agent run can use all the credit you have until the run ends or the credit reaches zero. A run only carries a ceiling if one is given when it starts: `/ultracode` accepts a maximum spend, and the run stops when it reaches it. Check your credit and your plan's run time before you start a long run. A monthly budget governor exists, but the operator sets it, not you: see [Budget governor](https://zerolog.ch/docs/governor).
choosing a model
## Check the price before you send
The picker shows the models available to your account and their current prices. Choose one that fits your budget. If a model is unavailable, follow the message in chat or pick another model from the list.
not a spend limit
## Hardware and abuse limits live elsewhere
This page is about money. How much CPU, memory and disk a container gets, how many run at once, and how long an agent may run are set by your [plan](https://zerolog.ch/docs/plans). A few actions, such as signing in, are separately rate-limited to slow down automated abuse; if one stops you, wait a moment and try again.
How unattended runs behave: [/goal and /ultracode](https://zerolog.ch/unattended-agents).
## What stops spending
Mermaid source of the diagram on this page:
```mermaid
%% Spend limits: what stops spending. Facts from server/billing.js, delegation.js, job-handlers.js (spendCapUsd), tiers.js goal_minutes.
flowchart TB
r["A paid model request"]
c{"Credit above zero?"}
stop["Paid models stop
nothing is billed later"]:::hot
d["Delegated sub-agent calls
at most 4 per turn,
default cap CHF 0.50 per turn"]
g["/goal run
ends at your plan's
time limit"]
u["Run started with a ceiling
(for example /ultracode)
stops at that ceiling"]
r --> c
c -->|no| stop
c -->|yes| ok["Request runs, still bounded by"]
ok --> d
ok --> g
ok --> u
classDef hot stroke:#8f7ac9
```
---
# Budget governor
The governor is a monthly guard that the operator of ZeroLog sets, not you. It adds up what the whole service spends on models this month. Once that total passes a threshold of the monthly budget, paid requests move to a cheaper model, so chats **keep going** instead of stopping outright. It is off until the operator turns it on and sets a monthly budget. The cheaper model is the one the operator named, or else the cheapest one available.
*The month, watched as a whole*
Past the monthly threshold
The model you picked
A cheaper model
- From then on, paid requests are answered by a cheaper model, and the chat carries on.
- Refusing the request. The downgrade happens instead of this, unless the operator has also switched on a hard stop at the full budget.
## Every downgrade leaves a line
The governor writes down each downgrade with the reason behind it. So when an answer comes back looking different from the one you expected, there is a line that explains it.
## What it does not do
It does not watch a single run or measure how fast a run spends, and you cannot set or see it from your account. Your own limit is your credit: at zero, paid models stop. See [Spend limits](https://zerolog.ch/docs/limits).
How unattended runs behave: [/goal and /ultracode](https://zerolog.ch/unattended-agents).
## What the budget governor does to a paid request
Mermaid source of the diagram on this page:
```mermaid
%% Budget governor: one paid request. Facts from server/governor.js (evaluate).
flowchart TB
r["A paid request
for the model you picked"]
h{"Hard stop on and
monthly budget used up?"}
refuse["Request refused"]
t{"Service-wide spend this month
past the threshold?"}
same["Answered by the model
you picked"]
down["Answered by a cheaper model
the chat carries on"]:::hot
log["Downgrade logged
with the reason"]
r --> h
h -->|yes| refuse
h -->|no| t
t -->|no| same
t -->|yes| down --> log
classDef hot stroke:#8f7ac9
```
---
# Plans
Not liable for Swiss VAT (Art. 10(2)(a) VAT Act); no VAT is charged.
Nicht mehrwertsteuerpflichtig (Art. 10 Abs. 2 lit. a MWSTG); es wird keine MwSt erhoben.
Free is an account. Mini, Pro, Max and Ultra add the machine.
## The five plans
Free has no container. Mini, Pro and Max have one container each, in growing sizes. Ultra has up to three that share one pool of vCPU, RAM and SSD. Terminal, code execution, builder, design engine, site builder, web editor and public ports all run in a container, so they need Mini or higher. Chat, files and the vault work on every plan. Pay yearly and you pay for ten months.
Autonomous runs (`/goal`) have a limit per run, not per month: a run may go on for 4 hours on Mini, 8 on Pro and 12 on Max, then it ends and you can start a new one. Only Ultra runs around the clock (24/7).
The plan decides the hardware. Credit pays for model usage and is a separate meter: topping up never changes your plan.
### Free
Per month **CHF 0**
- **Container**: None
- **Credit each month**: None included
- **Agent runs**: No
- **Deep Research**: No
- **Files**: 1 GB
- **Models**: Your own access or credit you buy
- **Support**: Help pages
### Mini
Per month **CHF 9**
Per year **CHF 90**
- **Container**: 1 · 1 vCPU · 4 GB RAM · 10 GB SSD
- **Credit each month**: CHF 2
- **Agent runs**: 4 h per run
- **Ports**: 1
- **Snapshots**: 1
- **Deep Research**: No
- **Support**: Email
### Pro
Per month **CHF 15**
Per year **CHF 150**
- **Container**: 1 · 2 vCPU · 8 GB RAM · 15 GB SSD
- **Credit each month**: CHF 4
- **Agent runs**: 8 h per run
- **Ports**: 1 + 1 domain
- **Snapshots**: 2
- **Deep Research**: Yes
- **Support**: Email
### Max
Per month **CHF 24**
Per year **CHF 240**
- **Container**: 1 · 4 vCPU · 12 GB RAM · 25 GB SSD
- **Credit each month**: CHF 8
- **Agent runs**: 12 h per run
- **Ports**: 2 + 1 domain
- **Snapshots**: 3
- **Deep Research**: Yes
- **Support**: Email
### Ultra
Per month **CHF 79**
Per year **CHF 790**
- **Containers**: Up to 3 · 12 vCPU · 48 GB RAM · 150 GB SSD in total
- **Credit each month**: CHF 30
- **Agent runs**: Unlimited (24/7)
- **Ports**: 8 + unlimited domains
- **Snapshots**: 20
- **Deep Research**: Yes
- **Support**: Email with priority
Ultra also has queue priority and guest VMs (VM class). Container links are 1 on Pro, 2 on Max and unlimited on Ultra; Mini has none. Scheduled runs and jobs are included from Mini on, and so is logging in your own AI subscriptions (CLI login) inside the container.
## Compute boost
On Pro, Max and Ultra you can add a boost: 2 vCPU, 8 GB RAM and 25 GB SSD for CHF 12 a month, up to two boosts per account, cancellable monthly. Mini has no boost, because it would be larger than the plan itself. On Ultra a boost counts against the whole pool.
## Credit: the models
Credit pays for any model in the catalogue at the provider's rate plus 2 %. You can buy it on every plan, Free included, as a monthly pack (a Stripe subscription that renews until you cancel) or as a one-time manual top-up:
- Credit S: CHF 5.90 adds CHF 5.50 of credit each month.
- Credit M: CHF 15.90 adds CHF 15.00 of credit each month.
- Credit L: CHF 51.50 adds CHF 50.00 of credit each month.
Pack credit is a monthly grant like the credit that comes with a plan: it is spent first and expires three months after it was added. Credit from a manual top-up does not expire.
A manual top-up has no ZeroLog fee and a 3 % currency buffer. It runs through MoonPay, with the minimum and checks shown at checkout. If you bring your own AI subscription or API key, that traffic does not use credit at all.
## Billing
- **Mini, Pro, Max and Ultra**: Billed **monthly or yearly through Stripe**. Yearly is ten months.
- **At checkout**: The figure shown at checkout is the authoritative one, because it is what your card is charged.
- **Changing plan**: You can change plan at any time. Stripe settles the part-month on your next invoice.
- **One at a time**: An account holds **one subscription**. Moving between plans is a change, never a second purchase.
- **Included credit**: Credit that comes with a plan is added each month after a successful invoice and is spent before credit you bought. It **expires three months** after it was added, cannot be paid out in cash and cannot be transferred. A yearly plan is billed once, but its credit still arrives month by month.
- **Cancel**: Credit you bought stays spendable. Your containers pause at the end of the paid period, and nothing is deleted.
How changes to the terms are announced, and with what notice, is set out in [Terms section 12](https://zerolog.ch/legal#tos-12).
## Two meters
### Plan
Buys the hardware and the tools on the plates above, and on every paid plan it also fills the credit meter beside it every month.
### Credit
Buys model usage. See [Credit](https://zerolog.ch/docs/credit).
The meters stay separate: hardware never becomes credit, and credit never becomes a bigger machine. Included credit is spent before credit you bought. See the live [pricing page](https://zerolog.ch/pricing) for current prices.
## Plans and credit are separate meters
Mermaid source of the diagram on this page:
```mermaid
%% Plans: two separate meters. Facts from server/tiers.js, plans.js, creditgrant.js, billing.
flowchart TB
plan["Plan: Mini, Pro, Max or Ultra
billed by Stripe, monthly or yearly"]
hw["Hardware and tools
container, ports, snapshots,
run time, Deep Research"]
pack["Credit pack S, M or L
Stripe subscription"]
top["Manual top-up
MoonPay, no expiry"]
inc["Plan and pack credit
added monthly, expires
after three months"]:::hot
bought["Topped-up credit"]
use["Model requests
spent in this order, stop at zero"]
plan --> hw
plan -->|monthly grant| inc
pack -->|monthly grant| inc
top --> bought
inc -->|spent first| use
bought -->|then| use
classDef hot stroke:#8f7ac9
```
---
# Two-factor authentication and passkeys
Sign-in runs through Keycloak at auth.zerolog.ch, ZeroLog's own sign-in service in Switzerland.
**Four factors** you choose
- **Password**: The classic. Optional once another factor is enabled.
- **Email**: A one-time code or a magic link sent to your verified address. Codes and links are kept in server memory only; ZeroLog sends the mail through SMTP2GO.
- **Authenticator app**: A six-digit TOTP code from any authenticator app.
- **Passkey**: WebAuthn: your device's own unlock (fingerprint, face or PIN).
## Set up a factor
Settings *→* Sign-in & security *→* Set up factors & email
1. *01*
The link opens your Keycloak account page.
2. *02*
Add an email address, an authenticator app or a passkey there.
3. *03*
Back in ZeroLog, the factor shows as **Ready**. A factor that is not set up stays greyed out and cannot be switched on. Email counts as ready once the address is verified.
## Choose the rules
Settings *→* Sign-in & security
Each factor has two switches. **Enabled** means it can be used to sign in. **Required** means it is asked for at every sign-in.
Pick a **preferred login** from the enabled factors; it is preselected at sign-in. Every sign-in asks for the method you choose plus all required factors.
To save, ZeroLog first asks you to confirm your sign-in (the button reads **Sign in to confirm**), then the button reads **Save rules**. The new rules apply from your next sign-in.
If you already had an authenticator app set up, your starting rules are password plus authenticator code, both required, until you change them.
## Passwordless
After you enable another factor, you can turn on **Passwordless**. The password is then no longer used to sign in.
## Passkeys
Nothing typeable is exchanged.
That is what makes a passkey resistant to phishing in a way a code is not.
Register more than one if you can: a laptop and a phone, say.
- Laptop
- Phone
A passkey lives on the device that made it, and a single passkey plus a lost device is a bad afternoon.
## From setting up a factor to the next sign-in
Mermaid source of the diagram on this page:
```mermaid
%% MFA: from setting up a factor to the next sign-in. Facts from server/account-security.js, server/routes/account-security.js, web account-security-panel.tsx.
flowchart TB
a["1. Set up a factor on the
Keycloak account page
(email, authenticator, passkey)"]
b["2. Factor shows Ready
in Sign-in & security"]
c["3. Per factor: switch
Enabled and Required
Pick a preferred login"]
p["Optional: Passwordless
(needs another factor)"]
d["4. Confirm your sign-in,
then Save rules"]
e["5. Rules stored with
your Keycloak account"]
f["6. Next sign-in: the method
you choose plus every
required factor"]:::hot
x["Not set up:
cannot be switched on"]
a --> b --> c --> d --> e --> f
c -.-> p
p -.-> d
a -.-> x
classDef hot stroke:#8f7ac9
```
---
# Secrets vault
Store credentials the AI needs (a deploy token, a database password) encrypted, and decide for each one what the AI may do with it.
## Three states per entry
**Visible:** the AI reads the value in plain text, but only in chats you release it for, and only when it asks. The value is encrypted in the app database, decrypted per chat on request, and every read is logged.
**Use only:** the AI sees a placeholder such as `zlph_github-token_` followed by an ID, and the list of hosts the value may be used for. The real value is stored in the [key service](https://zerolog.ch/docs/keyproxy) and inserted outside the app server and the box, only for an allowed host.
**Locked:** stored in the key service and never inserted. The AI sees and uses nothing.
Existing entries kept their meaning: those with model access became visible, all others locked and moved to the key service. Each value is encrypted with AES-256-GCM under a key the server or the key service holds, so this is not end-to-end encryption.
## How a use-only value is used
The box holds only the placeholder. A request leaves it through the egress broker to the key service, which checks that the ID is known, belongs to your account, is not locked and meets a host on the entry's **allowed for** list. Then it inserts the value in the URL, in headers and in small JSON, form or text bodies. In a command line, a placeholder stays a placeholder. A refused placeholder is not sent, and the AI is told why.
Use only works for HTTPS and Git targets, and for Postgres and MySQL or MariaDB. For databases the key service completes the login and requires TLS with a checked certificate. A target without TLS is refused, unless you mark that one entry **TLS not available**. SSH is not covered: pick visible or locked. Git no longer writes its token into the repository configuration; its credential helper hands the box a short-lived placeholder.
## Passkey confirmation
Adding a host to **allowed for**, switching an entry back to visible and showing a value need a fresh passkey confirmation, checked by the key service itself. It is tied to that one action on that one entry, valid for 60 seconds and usable once. The app only relays it.
A new passkey counts only after an existing one confirms it. Removing a passkey and resetting need a signature as well. If every passkey is lost, the key service emails a link to your account address and waits 72 hours after you confirm. After a first passkey or a reset, a waiting period of 24 and 48 hours blocks every confirmation.
The key service reports every change to your protection to you directly, by push message and email, without passing through the app: a new or removed passkey, a reset, a host added, a value shown.
## Visible entries and the chat
An entry can be attached to a chat only while its **Allow model access** switch is on. With the sandbox on, the value of a visible entry is written to a file under `/vault` in the container (mode 600). The model sees the name, user name, notes and path but not the value, and it can still read the file with commands, so a prompt-injected or misbehaving model could leak it. With the sandbox off, the server decrypts the value and includes it in plaintext in every request to the model provider for that chat.
A chat with an attached credential runs only on a route whose zero-data-retention commitment is confirmed. For Featherless that is the provider’s written commitment of 28 September 2026, which ZeroLog cannot enforce technically. On any other route the request is refused and nothing is sent. You can allow it manually for one chat; the chat then shows that the override is active, and you can revoke it.
Editing or deleting an entry, or detaching it from a chat, wipes the copy in that chat’s running sandbox. If the wipe fails, ZeroLog says so and asks you to delete that chat’s sandbox. A second net in front of every request to a model provider replaces the real values of your entries ([blackout](https://zerolog.ch/docs/redaction)).
Where this stops
### A passkey prompt does not show what you confirm
A taken-over app could still talk you into a confirmation, and you would see the message right after. Server tools such as the registrar borrow a short-lived placeholder or read one value for their own call, and every read is logged. Only you set the allowed hosts, not the AI. After a database login the data stream is not scanned for the value.
For keys that agents should use without ever seeing them, see the [key proxy](https://zerolog.ch/docs/keyproxy). Plain-language overview: [API key proxy for AI agents](https://zerolog.ch/api-key-proxy-for-ai-agents).
## Three states, and the path of a use-only value
Mermaid source of the diagram on this page:
```mermaid
%% Vault: three states and the path of a use-only value. Facts from server/vault.js, keyservice/src (tresor, einsetzen), egress broker.
flowchart TB
subgraph ST["One state per entry"]
direction LR
vis["Visible
the AI reads the value
in chats you release"]
use["Use only
the AI sees a placeholder
and the allowed hosts"]:::hot
lock["Locked
seen and used by nothing"]
end
box["1. Box
holds only the placeholder
zlph_name_ID"]
broker["2. Egress broker
passes the request on
with a signed preamble"]
ks["3. Key service (separate VM)
checks ID, owner, host list
inserts the real value"]:::hot
target["4. Target host from allowed for
HTTPS, Git, Postgres, MySQL"]
use --> box
box --> broker --> ks --> target
linkStyle 0 stroke:#8f7ac9
linkStyle 1 stroke:#8f7ac9
linkStyle 2 stroke:#8f7ac9
linkStyle 3 stroke:#8f7ac9
classDef hot stroke:#8f7ac9
```
---
# Blackout of secrets
> Secrets are blacked out before anything is stored or sent on.
**Secrets are blacked out before anything is stored or sent on.** Tool output, the agent log and requests to a model provider pass one blackout step.
- In tool output, three asterisks replace passwords inside addresses, proxy variables, values of names ending in secret or token, and unambiguous key formats. This runs in chat turns and in background job runs, before the output goes into the chat history or to the provider.
- Before a request goes to a model provider, the values of your [vault](https://zerolog.ch/docs/vault) entries are replaced as well, also as base64, hex or URL encoded text, by the placeholder or a short note.
- The command of an Ultracode agent shows in the log up to 2,000 characters, with the same patterns and your vault values blacked out.
- The [key service](https://zerolog.ch/docs/keyproxy) removes your key values from provider responses and swaps new secrets for placeholders.
## Where this stops
The patterns recognize formats and names. A secret in a format nobody could recognize, in a place the patterns do not cover, can still pass. Vault values are replaced only if you saved them. After a database login the data stream is not scanned for the value. For values you do not want the AI to read at all, set the entry to **use only** or **locked**.
## Where secrets are blacked out
Mermaid source of the diagram on this page:
```mermaid
%% Blackout of secrets: where values are replaced. Facts from server/ausgabe-schwaerzen.js, keyservice/src/schwaerzen.js, agent_log.
flowchart TB
out["Tool output
env, logs, command results"]
req["Request to a model provider
after the chat is assembled"]
log["Agent log
command up to 2000 characters"]
red["Blackout step
replaces secrets before
anything is stored or sent"]:::hot
db["Chat history and database"]
prov["Model provider"]
out --> red
req --> red
log --> red
red --> db
red --> prov
linkStyle 3 stroke:#8f7ac9
linkStyle 4 stroke:#8f7ac9
classDef hot stroke:#8f7ac9
```
---
# Key proxy for third-party keys
Store an API key for a third-party provider (Cloudflare, Tailscale, Stripe, and others) and agents in your chats can use it without ever seeing the real value. Only a placeholder reaches an agent; the real key stays in a separate, dedicated key service and is swapped in only on the way out to the provider.
## What to know
- You register a passkey before your first key — it's what guarantees that access can only ever be widened with your confirmation, never by the app alone.
- New keys start **read-only**. Pick a template (for example "DNS for one zone") to grant more, or switch on **full access** with a passkey confirmation. Outside full access, payments and orders always wait for your approval; under full access they go through without asking.
- You can bring your own AI provider key: 25 providers are listed, from OpenAI, Anthropic and Mistral to Swiss ones such as Infomaniak, and some are still marked beta or soon ([full list](https://zerolog.ch/docs/byok)). Requests with it go straight to that provider, so that provider's retention terms apply. ZeroLog does not guarantee zero data retention by third-party providers; Featherless's statement about Developer API content is a provider commitment, not a control ZeroLog can enforce per request.
- Any new secret a provider hands back — a freshly created token, for example — is caught and stored the same way. An agent only ever sees a placeholder for it too.
- Lock a key instantly from the panel. A locked key stops working everywhere at once.
- Containers with their own VPN access do not go through the key service and cannot use keys stored here — this is enforced technically, not just documented: such a workspace never receives key-service trust or secrets in the first place.
- A site hosted elsewhere reads its own code, including anything embedded in it — prefer a key limited to what that site actually needs.
Plain-language overview: [API key proxy for AI agents](https://zerolog.ch/api-key-proxy-for-ai-agents).
Each vault entry can also be set to **use only**: the AI sees a placeholder with an ID, and the key service inserts the value only for hosts you allow, now also for Postgres and MySQL. See [the vault](https://zerolog.ch/docs/vault) for the three states, the passkey confirmation and the path from box through broker and key service to the target, and [blackout](https://zerolog.ch/docs/redaction) for how values are removed from tool output and provider requests.
## Finding a provider
Open Settings, then Providers & keys. The page has two tabs: **AI providers** for your own AI key (see [your own API key](https://zerolog.ch/docs/byok)) and **Services & keys**, which is this key proxy. Each tab has its own address, `?tab=ki` and `?tab=dienste`.
- The catalogue holds over a hundred templates for DNS, hosting, mail, payments, databases, monitoring and more. Each template fixes the provider's host and how the key is sent, and sorts its actions into read and write. The exact number shows above the search box.
- Search by name, category or keyword. Press `/` to jump to the field, narrow by category, move through results with the arrow keys and use **Show more** for long lists. If nothing matches, the page offers to suggest the provider.
- About sixty services are listed as **Soon**, each with the reason it cannot yet be mapped to a fixed host and header.
## Microsoft 365
The Microsoft 365 (Graph) template covers the directory, mail, calendar, files, SharePoint pages and the list of Teams channels. The app signs in with client credentials: you enter the tenant ID or domain, the client ID and either a certificate (recommended) or a client secret. The key service trades them for short-lived Graph tokens by itself and talks only to `graph.microsoft.com`. The agent sees neither the secret nor the token.
- 25 actions. A new key starts read-only. The write template adds calendar entries, drafts, file contents and folders. Sending and replying to mail, and deleting mail, events or files, wait for your approval unless the key has full access.
- Only the global Microsoft cloud works; national clouds do not. Application permissions cover the whole tenant and need admin consent in Entra, so limit mailboxes and use Sites.Selected. Sending Teams messages and device-code sign-in are not available. The form lists the permissions each action group needs.
## What a hidden key does not prove
A placeholder prevents the supported agent route from returning the credential itself. Destination, method and path policies decide which operations are allowed. An allowed operation can still have harmful effects; narrow the provider permissions as well as the proxy policy. This is not protection against every compromise of the application or host. See [the request boundary](https://zerolog.ch/api-key-proxy-for-ai-agents) and [the separate vault](https://zerolog.ch/docs/vault).
## One request through the key proxy
Mermaid source of the diagram on this page:
```mermaid
%% Key proxy: one request from an agent to a provider. Facts from keyservice/src/pipeline.js, regeln.js, vorspann.js, ratenlimit.js, schwaerzen.js.
flowchart TB
a["1. Agent sends a request
with a placeholder
instead of the key"]
b["2. Egress broker hands
over with a signed
preamble (user, container)"]
subgraph KS["Key service (separate VM)"]
k1["3. Reject ambiguous
requests: path tricks,
duplicate headers"]
k2["4. Pick the profile by
host. Placeholder must
be yours, same container"]
k4["5. Classify the operation,
check the key's rule"]
k5["6. Rate brake
per risk class"]
k6["7. Insert the real key"]:::hot
k7["9. Redact your key values,
swap new secrets for
placeholders, send back"]
end
refused["403 refused"]
you["You approve unknown
operations, payments
(60 s, then 403)"]
p["8. Provider API
profile host only
no redirects"]
a --> b --> k1 --> k2 --> k4
k4 -->|allowed| k5
k4 -->|needs you| you
k4 -->|not allowed| refused
you -->|approved once| k5
k5 --> k6
k6 --> p
p --> k7
classDef hot stroke:#8f7ac9
```
---
# Your own API key
Already paying a provider? Add your key to use that provider account instead of your ZeroLog credit. The AI providers tab lists 25 providers: OpenRouter and Anthropic are live, 18 more are beta, 5 are marked soon.
## Where the charge lands
- **With your key** Your account at the provider
- **Without one** Your ZeroLog credit
## Your key stays private
Your account
Add the key in Settings → Providers & keys.
After saving
The full key is hidden.
Only the last four characters stay visible.
## What changes
An OpenRouter key replaces ZeroLog’s shared pool key for models routed through OpenRouter. An Anthropic key is checked with a live test call when you save it. If Anthropic accepts it, Claude models appear in your model list marked *(your key)*. A request on an Anthropic model with your key never falls back to ZeroLog credit or to the pool; if the key stops working you get an error instead.
ZeroLog books no cost for answers that use your key, so no per-answer price appears. The charge and the exact price are in your provider account. Your plan still decides hardware such as containers; the key only replaces credit. Saving a key is limited to five attempts per minute, and the BYOK permission must be on for your account (new accounts have it). If an admin withdraws it, your stored keys stay but are not used.
## The providers
Live: OpenRouter, Anthropic. Beta: OpenAI, Google (AI Studio), Mistral AI, xAI, DeepSeek, Groq, Together AI, Fireworks AI, Cohere, Cerebras, DeepInfra, Hugging Face, Moonshot (Kimi), Alibaba Qwen, Featherless, Infomaniak AI Tools, Phoeniqs, Scaleway. Soon: Perplexity, Azure OpenAI, AWS Bedrock, Safe Swiss Cloud, Prem AI. Infomaniak AI Tools and Phoeniqs (both Switzerland) are beta; Safe Swiss Cloud and Prem AI are marked soon.
## Beta providers
Beta means the connection exists and passes tests against a simulated provider, but it has not been verified with live traffic yet. Providers marked soon cannot be saved.
## Providers outside Switzerland and the EU
DeepSeek (servers in China), Moonshot Kimi (a Chinese company, servers in Singapore) and Alibaba Qwen (Singapore) show a notice before you can save a key. It says that your requests and the answers go straight to the provider under its terms and local law, that Switzerland and the EU do not recognise an adequate level of data protection there and authorities can access the data, and that you should not send confidential content or other people’s personal data. It links to the provider’s privacy policy. You must tick *I have read this and still want to use this provider*. The server refuses to save the key without it. The confirmation is stored with its date for you and that provider, so changing the key later needs no second tick.
## Helper and checking models
The small jobs around a chat (summaries, /goal checks, Deep Research helpers, the Ultracode planner and its checkers, and more) run on a small model of your provider, on your key, without ZeroLog credit. Anthropic uses Claude Haiku 5.5 by default; for the other providers ZeroLog picks the smallest model of the newest family your key lists. You can choose another per provider under Settings, Advanced. On the platform these jobs stay on Featherless. See [helper and checking models](https://zerolog.ch/docs/helper-models).
Without your own key, models are paid from ZeroLog credit. Packs and plan prices: [Pricing](https://zerolog.ch/pricing).
## Which account pays for a request
Mermaid source of the diagram on this page:
```mermaid
%% BYOK: which account pays for a request. Facts from server/provider-keys.js, routes/provider-keys.js, byok-anthropic.js, chatusage.js.
flowchart TB
k["1. You add a key in
Settings, Providers & keys
(OpenRouter or Anthropic)"]
st["2. Stored encrypted
only the last four
characters stay visible"]
pr["Anthropic: live test call
when you save"]
q["3. Chat request:
valid key and BYOK
permission?"]
y["Your provider account
pays and prices it"]:::hot
n["ZeroLog credit pays"]
b["ZeroLog books no cost
for the answer"]
k --> st --> q
k -.-> pr -.-> st
q -->|yes| y
q -->|no| n
y --> b
classDef hot stroke:#8f7ac9
```
---
# Helper and checking models
> With your own key, the small jobs around a chat run on your provider.
**With your own key, the small jobs around a chat run on your provider.** Before, several of them quietly switched to a platform model or failed.
- Helper jobs are summaries of long chats, [/goal](https://zerolog.ch/docs/goal) checks, Deep Research helpers, memory suggestions, the prompt improver, the design loop, the site builder, the Ultracode planner and reducer, the tool checker, second opinions from another model, Buddy, legal questions, Smart Auto and the small model of the terminal agent.
- On your own key they run on a helper model of that provider. Your provider bills them, so they need no ZeroLog credit and a balance of zero does not stop them.
- The default is a small, fast model from the list your key offers. For Anthropic it is Claude Haiku 5.5. For the others ZeroLog picks the smallest model of the newest family with names such as nano, mini, flash, small or haiku. If your key lists no small model, the run model does the job itself.
- You can choose another model per connected provider under Settings, Advanced, **Helper model with your own key**. The model that describes pictures for text-only models is set there too.
- A chat on a platform model keeps its Featherless helper. The penetration test keeps its fixed Featherless list on purpose.
## The checking model of an Ultracode or goal run
- On your own key the checker can be any model on your keys or a Featherless model. On the platform it must be a Featherless model. Models of other platform providers are refused.
- The check happens at the start, also for the price estimate. A checker that is not allowed is refused at once, and the answer lists up to eight allowed models, the closest first.
- A checker on your own key counts as zero platform cost.
We tested this with a replica of the provider lists. We have not yet run every helper against a real key of each of the 25 providers, so the default for providers other than Anthropic depends on what your key lists.
See [your own API key](https://zerolog.ch/docs/byok) and [/ultracode](https://zerolog.ch/docs/ultracode).
## Which model does the helper jobs
Mermaid source of the diagram on this page:
```mermaid
%% Helper and checking models: which model does the small jobs. Facts from server/pruefmodell.js (hilfsmodellFuer, PRUEF_STANDARD, pruefModellErlaubt).
flowchart TB
job["A chat starts a helper job
summary, /goal check, planner,
checker, memory, second opinion"]
own["Chat on your own key"]
plat["Chat on a platform model"]
set["Your setting for that provider
Settings, Advanced"]
def["No setting: a small model from
your key's list
Anthropic: Claude Haiku 5.5"]:::hot
run["Runs on your key
billed by the provider
no ZeroLog credit"]
fe["Featherless helper
as before"]
job --> own
job --> plat
own --> set
set -->|"chosen"| run
set -->|"not set"| def
def --> run
plat --> fe
linkStyle 4 stroke:#8f7ac9
linkStyle 5 stroke:#8f7ac9
classDef hot stroke:#8f7ac9
```
---
# Your own vendor subscription
Sign a CLI in with your own account and it runs on your subscription instead of your credit here.
## The sign-in
Select a private workspace you own. For native Claude Code, open CMD in your chat, choose *Your account* and run `/login`. For Codex, use **Connect account** under **AI accounts** in settings. This needs a container, which the Free plan does not include.
1. The CLI starts /login or device sign-in
2. you open it
3. the CLI confirms completion
***Nothing is intercepted or replayed.** The path leaves the CLI and comes back to it, because sign-in mirrors each CLI's own flow: it is that CLI doing the signing in.*
## Once signed in
- **The credential**: Available only in a workspace owned by you that nobody else can write to.
- **The login**: It is temporary. After the runtime stops or is recreated, connect again. Sign-in is refused in a shared workspace, because anyone who can write there could read the token.
- **The bill**: That vendor bills you on your subscription. **Nothing reaches your credit ledger.**
- **The limits**: Your subscription's own limits apply unchanged. This service **does not raise them, extend them, or reroute a request that has hit one**.
## What does not reach here
- **No automatic model switch.** Customer catalogue requests show an error when a model is unavailable; choose another model yourself. Vendor-account requests are not transferred to another provider either.
- **Spend limits and the budget governor**, because there is no spend of ours to cap.
## Different native sign-ins
In Claude Code, use `/login`. For Codex, run `codex login --device-auth`, open the displayed URL and enter the device code on OpenAI’s page. Codex waits for completion itself: no code is pasted back into ZeroLog.
Tool conversations and model settings stay separate. Files, packages and caches can remain in one workspace. Give the next tool a handover with the task, changed files and tests. [See the tool-switch workflow](https://zerolog.ch/bring-your-own-ai-subscription).
## Signing a CLI in with your own subscription
Mermaid source of the diagram on this page:
```mermaid
%% Own vendor subscription: native CLI sign-in and who bills. Facts from server/agent.js, web accounts-panel.tsx, plans.tsx.
flowchart TB
w["1. Pick a private workspace
you own (needs a container;
shared workspace is refused)"]
c["2. Start the CLI sign-in
Claude Code: /login in CMD
Codex: device sign-in"]:::hot
v["3. You approve on the
vendor's own page"]
d["4. The CLI confirms
itself, nothing is pasted
into ZeroLog"]
r["5. Work runs on your
vendor subscription"]
l["Vendor limit:
error, no reroute"]
t["Runtime recreated:
connect again"]
w --> c --> v --> d --> r
r --> l
r -.-> t
t -.-> c
classDef hot stroke:#8f7ac9
```
---
# Anonymous chats
An anonymous chat keeps its saved messages private to you. The model provider still receives your prompt so it can answer.
In your browser
Can I hand in my notice before the fixed term ends?
Private to you
Your saved conversation
with your passphrase
*Unlock the conversation with its passphrase to read saved messages.*
## Who else can read it
what they can see
- **Whoever runs the service**
People operating the service cannot read your saved messages.
- **The model provider**
The prompt still reaches the provider to be answered.
Your prompt is still sent to the selected model provider to generate an answer. Anonymous turns ask for zero-data-retention routing and fail instead of using a provider that cannot offer it. Whether a route enforces that depends on the provider; see [Zero data retention](https://zerolog.ch/docs/zdr).
## How it works
Open the chat’s settings and choose **Make anonymous**. If the chat already has messages stored in plain text, ZeroLog refuses until you choose to delete that plain text. Your browser derives a key from your passphrase and the chat’s public salt (PBKDF2-SHA256, 600,000 rounds) and encrypts each message with AES-GCM. The server stores only the ciphertext, in a storage area of its own, plus token and cost counters. It holds no key and has no way to decrypt.
Each turn, your browser sends the decrypted history along with your new message, because the server cannot rebuild it. The reply streams back and your browser encrypts the pair before saving it.
An anonymous chat is plain conversation. It is left out of search, memory and compaction, and you cannot set a goal on it.
## Storage and deletion
If ZeroLog cannot save a message, you will see an error. It is not saved as a regular chat. Deleting an anonymous chat also deletes its saved messages. **Turn off anonymous** destroys the saved ciphertext too, because the server could not convert it back.
## The passphrase
Lose the passphrase and it is gone permanently.
No support request can recover it. Your browser does not store it, so after a reload or a closed window you enter it again.
## Every new chat
An account preference, **anon_by_default**, makes every new chat start anonymous. The current web app has no switch for it in its settings screens yet. While it is on, the browser extension will not start chats and points you to the app.
## Where an anonymous chat is encrypted
Mermaid source of the diagram on this page:
```mermaid
%% Anonymous chats: what is encrypted where. Facts from server/compact.js, server/anon-store.js, web lib/anon.ts.
flowchart TB
a["1. Your browser derives a key
passphrase + public salt
never stored"]
b["2. Browser encrypts each
message (AES-GCM)"]
c["3. Each turn: browser sends
decrypted history
+ new message"]
d["4. Server forwards the prompt
asks for zero data retention"]
e["Model provider
sees the prompt to answer"]
f["5. Reply streams back,
browser encrypts it"]
g["Server storage: ciphertext
plus token and cost counters
no key, cannot decrypt"]:::hot
a --> b --> c --> d --> e --> f
f -->|ciphertext only| g
classDef hot stroke:#8f7ac9
```
---
# Zero data retention
Retention depends on the model used for each answer
ZeroLog shows whether zero data retention was enforced for an answer, stated generally by the provider, or not guaranteed.
## Where it stands today
*Check what applies to each answer*
Your conversation **Status shown**
The conversation shows the retention status for the model that answered.
After the answer **Review**
The retention status appears with the answer. Review it before sharing sensitive information.
Provider statements
A provider’s general statement does not confirm how a particular request was handled. Check the status in the chat transcript.
## The three statuses
**Zero data retention:** the request asks for it explicitly and the recipient enforces it. This applies to models served through the OpenRouter pool. If no provider can meet the request, the answer fails with “No zero-data-retention provider available for this model” instead of going elsewhere.
**Provider statement:** Featherless states generally that it does not store prompts or replies. No retention control is sent with each request, so this is the provider’s statement and not something ZeroLog enforces.
**No statement recorded:** for every other route, ZeroLog has no non-retention statement on file. The chat says so.
With an automatic model choice, the status is the one for the model that answered. [Anonymous chats](https://zerolog.ch/docs/anon) and chats with credentials both ask for zero-data-retention routing.
The chat transcript identifies the provider and its retention status for the route that answered.
A chat that holds credentials is stricter: it is refused, without any network call, on a route whose zero-data-retention commitment is not confirmed, and it never falls back to another provider. Featherless counts as confirmed through its written commitment of 28 September 2026, which is the provider’s statement and not enforced by ZeroLog. You can allow such a chat manually for one chat; the chat then shows that the override is active.
## Retention status and the credentials gate
Mermaid source of the diagram on this page:
```mermaid
%% Zero data retention: three route states and the credentials gate. Facts from server/retention-core.js, retention.js, routes/chats.js, web retention-note.tsx.
flowchart TB
q["Answer is routed
to a model"]
e["Zero data retention
request asks, recipient
enforces (OpenRouter pool)"]
p["Provider statement
Featherless says it does
not store; not enforced"]
n["No statement recorded
all other routes"]
s["Status shown
with the answer"]
g["Chat holds credentials?"]:::hot
ok["Runs"]
rf["Refused, nothing sent
unless you allow it
for this chat"]
nf["No provider can meet it:
error, no fallback"]
q --> e
q --> p
q --> n
e --> s
p --> s
n --> s
e -.->|none available| nf
s --> g
g -->|enforced, or provider
statement on file| ok
g -->|no statement| rf
classDef hot stroke:#8f7ac9
```
---
# Image generation
Describe a picture and get one back. One image engine, three formats, a fixed price per image.
## Make an image
Open **Images** in the sidebar. Write what you want (up to 400 characters), pick a format and press **Generate image**. The button shows the price in CHF. It comes off your ZeroLog balance when the image is done. If generation fails, nothing is charged.
There is no model to choose. Every image comes from one engine, Flux, run by Infomaniak, and your description is sent to them. An image can't be edited: each request makes a new one.
***Square** The default shape.*
***Wide** Banners and backgrounds.*
***Tall** 1024 × 1792.*
## Keep it
- **Download** the file to your device.
- **To container** copies it into the container of one of your chats.
- **Delete** it whenever you like. Images are removed after 30 days anyway.
## Let the AI make images
Off by default. Switch on **Allow AI to generate and inspect images** on the Images page, or type `/image on` (`/image off` turns it off). The AI can then generate an image, look at it and copy it into the chat's container. Each image costs the same fixed price and counts against the chat's spending limit.
Limits: one image at a time per account, six per minute, and a daily cap for the whole service.
## What it costs
Each image has a fixed price, shown before you press the button. See [Cost](https://zerolog.ch/docs/cost).
## How an image request runs
Mermaid source of the diagram on this page:
```mermaid
%% Image generation: from a request to a stored image. Facts from server/images-core.js (take, start), images-router.js, delegation.js.
flowchart TB
a["You: Images page
description up to
400 characters, one format"]
b["The AI in a chat:
generate_image, only if
you switched it on"]
subgraph SV["ZeroLog server"]
c["1. Check the price you saw
is current and your balance
covers it"]
d["2. Reserve the price
on your balance
(held, not charged yet)"]
e["3. One engine makes
the image: Flux via
Infomaniak, one job
per account"]:::hot
f["4. Charge the price,
store the image 30 days"]
end
x["Refused, nothing charged
(price changed, balance
too low, too many requests)"]
r["Reservation released,
nothing charged"]
g["Download"]
h["To container"]
i["Delete any time"]
a --> c
b --> c
c -->|ok| d
c -->|not ok| x
d --> e
e -->|failed| r
e -->|done| f
f --> g
f --> h
f --> i
classDef hot stroke:#8f7ac9
```
---
# Gallery and backgrounds
The gallery collects what the AI made for you in chats. Your background is set in Settings, from a file on your device.
## What the gallery shows
The Gallery page, headed Media, has two parts.
- **Designs**: Every page, component, logo, diagram, animation or 3D design the AI wrote in your chats, with its version. *Open design* reopens it; the chat link jumps to where it was made. See [Design engine](https://zerolog.ch/docs/design-engine).
- **Other media**: Video, audio and other files a connector saved for you, with prompt, date and a download link. Filter by provider and kind. The list is empty if you have none.
Images you generate are not here. They live on the Images page, see [Image generation](https://zerolog.ch/docs/images).
## Backgrounds
Settings, Appearance, Background. It sits behind the workspace; text stays readable on its surfaces.
- **Type**: None, colour, gradient, one of six patterns, your own image, or an animated background (particles, gradient flow, matrix).
- **Own image**: PNG, JPG or WebP up to 3 MB. ZeroLog keeps one re-encoded copy per account, at most 2560 pixels wide, without location and camera data. *Remove image* deletes it from the server.
- **Blur and dim**: Two sliders soften and darken any background.
- **Plans**: Own images and animated backgrounds need Mini or above.
To use a picture you generated, download it from the Images page, then upload it here.
## Where your images, designs and background live
Mermaid source of the diagram on this page:
```mermaid
%% Gallery and backgrounds: where things live. Facts from the Gallery page, the Images page, Settings > Appearance, server/routes/ui-prefs.js.
flowchart TB
subgraph IM["Images page"]
gen["Images you generate
stored 30 days"]
end
subgraph GA["Gallery page (Media)"]
des["Designs the AI wrote
pages, logos, components,
diagrams, animations, 3D"]
med["Other media from
connectors: video, audio"]
end
dev["A picture file
on your device"]
subgraph ST["Settings, Appearance, Background"]
bg["None, colour, gradient,
pattern, animated, or your
own image (PNG, JPG, WebP,
up to 3 MB)"]:::hot
end
gen -->|Download| dev
dev -->|upload, Mini and up| bg
classDef hot stroke:#8f7ac9
```
---
# Design engine
The AI in your chat designs websites, logos and banners itself and uses the design engine to check its work.
With `/design`, the AI in your chat does the design work itself. It writes the layout, text and code, and uses the design engine as its toolkit to build, look at and check what it made. Nothing is handed to a separate background model.
Type `/design` followed by what you want, for example `/design logo for a joinery in Baden` or `/design website for a café, three pages`. A preview opens on the right of the chat as soon as the AI starts, and it shows the current state after every build. A logo, banner or single page is drawn live while the AI writes it, then saved to the chat's container.
## What the toolkit does
| Step | What happens |
| --- | --- |
| Project | Sets up a website project in your chat's container: a checked starting structure, your colours and fonts as design tokens, and a start page for each page you asked for. Colour pairs that fall below WCAG AA contrast are raised automatically. |
| Build | Installs, type-checks, lints and builds the site, then starts it and opens every page. Problems come back with the file and line, so the AI can fix them and build again. |
| Look | Opens the running site in a real browser at 360, 768 and 1440 px wide, measures overflow, contrast, line length, tap targets and layout, and hands the screenshots to the AI to judge. |
| Security | Checks the code and dependencies for known risks, the required security headers, leaked secrets and, if the site has a backend, who can reach which data. |
| Export | Packages the finished site for Docker, static hosting, Vercel, Netlify, Cloudflare Pages, a web hosting folder or ZeroLog, and starts each package once to make sure it runs. It runs only after the build and the security check are green. |
| Render | For single pieces such as a logo or banner: renders an SVG or HTML file offline and returns a screenshot with measurements. For a CSS animation it returns one sheet with up to four moments side by side. |
| Recipes | Before building motion, an animated icon or an image layout, the AI fetches a short card of tested values. A video or clip export is a Remotion project in your container. |
## Designs as files you can change
Any AI model you choose in ZeroLog can design. Every SVG, HTML, JSX or TSX file the AI writes in your workspace becomes a design in the panel and renders at once. Ask for logos, icons, interface mockups, landing pages, diagrams, animations or React components.
- **Preview, Code, Versions, Files.** Each file keeps its history: every change is a new version you can view or restore. Variants are separate files side by side.
- **Open any design again.** Under each file the AI wrote, *Preview* and *Code* reopen the panel, even after you closed it. On the *Files* page, click a file to see it rendered or as source; React files render there too.
- **Phone, tablet, desktop.** Switch the preview width with one click.
- **Point at what to change.** Choose *Select element*, click a part of the design and write what should change. The AI gets the file, the line and the element, and changes only that part.
- **React without setup.** A component with `export default` renders directly. React is built in; other packages show a placeholder in the preview.
- **No network in the preview.** Designs run sealed off: they cannot load anything from the internet and cannot reach your ZeroLog session. Fonts and images belong in the file itself.
## Good to know
- **It is enabled per account.** Without it, `/design` has no toolkit to call.
- **Your chat needs a container.** The work happens in it, so you can open, change and download every file.
- **The AI judges the screenshots itself.** That needs a model that can read images. With a text-only model the AI gets the measurements, but not the pictures.
- **No model costs for the tools.** Building, looking and checking run in your container and on our servers. You pay only for the AI's own messages, as in any chat.
- **Code only.** Designs use HTML, CSS and inline SVG. The design engine does not generate images or reuse images from your gallery.
- **Fonts are served locally.** No font is loaded from a third-party service, so visitors' addresses are not passed on.
For a website that should be built without you watching, `/site` still starts the unattended builder. See Website builder.
## From /design to a checked result
Mermaid source of the diagram on this page:
```mermaid
%% Design engine: /design to a checked result. Facts from server/werkzeuge/designkasten.js (design.projekt, bauen, ansehen, sicherheit, exportieren, rendern, rezept), site-builder/werkzeugkasten-regeln.js (BREITEN, EXPORTZIELE), web/apps/web/lib/slash.ts.
flowchart TB
a["1. You type /design
and what you want"]
b["2. The AI in your chat designs it
itself (needs a container)"]
subgraph TK["Toolkit (your container and the server's browser, no model cost)"]
r["Render: logo, banner or single
page, offline, screenshot
and measurements"]
p["Project: structure, tokens,
a start page per page"]
bu["Build: install, type-check,
lint, build, open every page"]
lo["Look: real browser at 360, 768
and 1440 px, screenshots
and measurements"]:::hot
se["Security: code, dependencies,
headers, secrets, data access"]
ex["Export: Docker, static, Vercel,
Netlify, Cloudflare Pages,
webspace, ZeroLog"]
end
pv["Preview panel: Preview, Code,
Versions, Files"]
a --> b
b -->|single piece| r
b -->|website| p
p --> bu
bu -->|errors with file and line| b
bu --> lo
lo -->|the AI judges the screenshots,
fixes and builds again| bu
lo --> se
se --> ex
r --> pv
ex --> pv
classDef hot stroke:#8f7ac9
```
---
# Connectors
Connect Gmail, Google Drive or Google Calendar to your account. It is optional: you choose whether to connect, and you can disconnect at any time.
No connector
ZeroLog works fully without one.
## You connect it
Connecting is your action, with your own account. A Google connection asks for read access first. Write access is a separate consent.
## While it is connected
Every call goes through ZeroLog's egress broker, which allows only the hosts listed for the connector, and is logged under Activity.
- Available only where ZeroLog has enabled it for your account.
- Even with write access, each kind of action is granted separately and confirmed the first time.
## You disconnect it
Disconnect removes the saved connection, the activity log and the media from that connector. For Google, the access is also revoked at Google and the indexed copy is deleted.
No connector
You can connect the service again whenever you choose.
## One connector call, and the way out
Mermaid source of the diagram on this page:
```mermaid
%% Connectors: one call, and the way out. Facts from server/connectors.js (connectorFetch, record, disconnect), server/google.js (read scopes first, write separate, revoke + purge), web connectors-panel.tsx.
flowchart TB
a["1. You connect your own
account. Google asks for
read access first"]
b["2. In a chat, you or the AI
ask for something from the service"]
subgraph ZL["ZeroLog"]
c["3. Egress broker:
only hosts listed
for the connector"]:::hot
w["Write needs the separate write
consent, granted per action,
confirmed the first time"]
d["4. Call logged under Activity"]
end
g["5. The service (Google)"]
no["Refused"]
x["You disconnect: access revoked
at Google; connection, log,
media and indexed copy deleted"]
a --> b --> c
c -->|host not listed| no
c -->|read| d
c -->|write| w
w -->|allowed| d
w -->|no consent| no
d --> g
g -.-> x
classDef hot stroke:#8f7ac9
```
---
# Machine-readable versions
Use ZeroLog’s Markdown pages to read documentation in an AI tool or fetch it from a script.
## Find a Markdown page
Open [`/llms.txt`](https://zerolog.ch/llms.txt) (also at `/.well-known/llms.txt`) for an overview and links to the available Markdown pages. Documentation topics have a `.md` version, such as [`/docs/start.md`](https://zerolog.ch/docs/start.md); the topic index is [`/docs.md`](https://zerolog.ch/docs.md). Add `.md` to any other supported English page: the home page is `/index.md`, pricing is [`/pricing.md`](https://zerolog.ch/pricing.md), and the product pages, `/agents.md`, `/faq.md` and `/investors.md` work the same way. News is at `/news.md` and `/news/.md`. In German, only `/de.md` and `/de/agents.md` exist.
The overview, the home page, the agents page, the FAQ, the pricing and every documentation topic are in one file: [`/llms-full.txt`](https://zerolog.ch/llms-full.txt).
## Request Markdown at the usual URL
On a page with a Markdown version, send `Accept: text/markdown` to request Markdown at its usual URL. For example, request `/docs/start` with this header. The response carries `Vary: Accept`. Supported pages include the homepage, pricing, documentation and published news. Pages without a Markdown version, including `/status`, `/extension`, legal pages and German pages other than the two above, do not support this negotiation.
If the response is HTML, use the explicit `.md` address linked from `/llms.txt`. A cached response may not reflect the format you requested.
## Hints in the HTML response
Pages with a Markdown version answer with a `Link` header that names the `.md` address (`rel="alternate"`, `type="text/markdown"`) and points to `/llms.txt` (`rel="describedby"`). `/status` and `/extension` point to `/llms.txt` only.
## Ways to read ZeroLog as Markdown
Mermaid source of the diagram on this page:
```mermaid
%% Machine-readable versions: three ways in. Facts from server/maschinenlesbar.js (llmsTxt, zuordnung, zuordnungMd, markdownSenden, LLMS_VERWEIS).
flowchart TB
t["Your AI tool
or script"]
l["GET /llms.txt or
/llms-full.txt: overview,
links, whole docs in one file"]
m["GET a page with .md
for example /docs/start.md"]
n["GET the usual URL
with Accept:
text/markdown"]
q["Does the page have
a Markdown version?"]
md["Markdown
Link header names
the HTML page"]:::hot
html["HTML
Link header names the
.md page and /llms.txt"]
t --> l
t --> m
t --> n
l -.->|links to| m
m -->|supported page| md
n --> q
q -->|yes| md
q -->|no| html
classDef hot stroke:#8f7ac9
```
---
Browser push
# Push notifications
An approval is waiting, or a run stopped while you were away. Push tells you, so you can stop watching the tab.
## When it tells you
You Desktop · Android
- **An approval is waiting**
- **A run stops without you**
The work
*Both of these happen while you are looking at something else. Without push you find them by going back to the tab. A run that stops means a paused or failed job, or a reply that ended while no stream was attached.*
## Allowing it
Allow notifications in your browser to receive browser alerts on desktop or Android. In the ZeroLog Android app, subscribe to push notifications and allow them in your device settings. Legal and Privacy alerts are available only through the app subscription.
In the ZeroLog Android app, open Settings, then Push notifications. Turning push off there turns it off for your whole account, on every device. Opening or updating the app will not turn it back on. Switch it on again in the same settings screen when you want notifications.
Allow notifications for ZeroLog in Android settings. Battery optimization can delay notifications. Android may run its background checks later than scheduled.
Read the [Android app update](https://zerolog.ch/news/free-page-block-typescript-check) in the newsroom.
Each of the two moments has a page of its own: [Command approval](https://zerolog.ch/docs/approval) and [Jobs and logs](https://zerolog.ch/docs/jobs).
## Other alerts
Choose which alerts you want under Settings, then Notifications. Alerts that wait for you are on by default. A job that finished, a scheduled task that ran and the daily digest are off until you switch them on.
- An approval is waiting, or a job paused or failed. These alerts are on by default.
- A reply finished or failed while you were away. These alerts are on by default and are sent only when no live stream is attached. They never include chat content. App notifications include the chat number; browser notifications do not.
- Your credit fell below its low-balance threshold. The default threshold is USD 1. This alert is on by default and is sent once each time your balance crosses below it.
- A sign-in came from a new location. This alert is on by default.
- A message from your assistant: the AI can send you one short line itself, for example when long work finished. Off by default; at most 5 per hour and 20 per day, never with credentials, code or file content. The app shows a fixed text with a link to the chat.
- A provider key needs your approval, was switched to full access, was locked or deleted, or was used in an unexpected way. These alerts are on by default and never name the provider.
- We published a Legal or Privacy newsroom update. This alert is off by default and is available through app notifications, not browser push.
App messages use fixed text. For each kind of alert the app gets at most one message a minute, and at most 20 an hour in total.
Read the [newsroom update](https://zerolog.ch/news/notifications-replies-credit-sign-ins) about these notification changes.
## From an event to your device
Mermaid source of the diagram on this page:
```mermaid
%% Push: from an event to your device. Facts from server/push.js (EVENTS, prefsFor, notify, sendePush, DROSSEL).
flowchart TB
ev1["An approval is waiting
for you"]:::hot
ev2["Job paused or failed,
reply finished while away,
credit low, new sign-in"]
ev3["Job finished,
scheduled task ran,
Legal or Privacy update"]
pref{"Is the switch on?
Settings, Notifications"}
off["Nothing is sent"]
subgraph CH["Two channels, independent"]
b["Browser push to each
device you turned on"]
a["ZeroLog Android app
via its own ntfy topic.
Fixed text, no chat content.
1 per kind per minute,
20 per hour"]
end
you["Your device"]
ev1 --> pref
ev2 --> pref
ev3 -->|off by default| pref
pref -->|off| off
pref -->|on| b
pref -->|on| a
b --> you
a --> you
classDef hot stroke:#8f7ac9
```
---
One boundary, two directions
# Webhooks
Let an outside system trigger work here, or be told when something finished.
Outside
ZeroLog
From *· Outside*
An outside system sends a request.
Inbound
Accepted
To *· ZeroLog*
The request triggers work here.
From *· ZeroLog*
Something here finished.
Outbound
To *· Outside*
The outside system is told.
## Only authorised requests are accepted
Unauthorised incoming requests are rejected before they can trigger work. The token in the URL is the credential. An unknown or revoked token gets a 404, so a guess reveals nothing.
The square on the inbound path above
## The three kinds
### Wake-up webhook · inbound
An outside system such as CI, GitHub Actions or Zapier sends `POST /api/hook/`. ZeroLog starts a [/goal](https://zerolog.ch/docs/goal) job in the chat and container you attached to the webhook, and answers 202 with the job number.
- Model, objective, completion criteria and spend cap are frozen in the webhook. By default the caller cannot change them. If you allow it, the caller can append free text of up to 4,000 characters.
- A paused container is woken first, unless you turned that off.
- Limits per webhook: 6 triggers a minute and 100 a day by default, up to 60 and 1,000. Beyond that the answer is 429. With no prepaid credit left the answer is 402. Each account can have 3 active wake-up webhooks.
- The token is shown once when you create the webhook and is stored only as a hash. A browser visit (GET) is refused with 405.
### Event webhook · inbound
Real events from Stripe, GitHub or another service are posted to `POST /api/inject/`. ZeroLog appends each one as a line of JSON to `/work/.webhooks/incoming.jsonl` in the container, so the agent can build and test against actual payloads.
- Each line holds the time, method, headers and body. Authorization, cookie and API-key headers are replaced before anything is written.
- Bodies up to 256 kB are accepted. The default limit is 60 events a minute, up to 120.
- This kind has no settings screen yet. You create and revoke it through the API (`/api/injects`).
### Decision webhook · outbound
You give ZeroLog a URL, an optional secret and the events you care about: `goal_set`, `turn_end`, `needs_approval`, `risk_check`, `goal_done` and `tool_call`. When one happens, ZeroLog sends a signed POST and acts on your reply: `continue`, `stop`, `approve`, `deny` or `ask`, with an optional instruction and reason.
- With a secret, each request carries an `X-ZeroLog-Signature` header with a timestamp and an HMAC-SHA256 of the body. Receivers should reject timestamps older than five minutes.
- The timeout is 5 seconds by default, between 1 and 30. ZeroLog retries once.
- If your service fails, times out or is not an allowed destination, ZeroLog's built-in evaluator decides instead. Redirects are not followed.
Manage the wake-up and decision kinds under Settings, Webhooks.
## The three kinds of webhook
Mermaid source of the diagram on this page:
```mermaid
%% Webhooks: the three kinds and what each checks. Facts from server/hooks.js, webhookinject.js, webhookai.js, routes/hook.js, routes/inject.js.
flowchart TB
subgraph OUT["Decision webhook (outbound)"]
direction TB
o1["Event in your chat:
approval, risk, tool call"]
o2["Signed POST to
your URL"]
o3["Your reply: continue,
stop, approve, deny or ask"]
o4["No answer in time:
built-in evaluator decides"]
o1 --> o2
o2 -->|answered| o3
o2 -->|failed or blocked| o4
end
subgraph EVT["Event webhook (inbound)"]
direction TB
e1["Service sends
POST /api/inject/token"]
e2["Same token and
pair checks, rate limit"]
e3["Append one JSON line to
incoming.jsonl in
the container"]
e1 --> e2 --> e3
end
subgraph WAKE["Wake-up webhook (inbound)"]
direction TB
w1["CI or Zapier sends
POST /api/hook/token"]
w2["Token known? Chat and
container still yours?"]:::hot
w3["Rate limits and
prepaid credit"]
w4["Wake container,
start frozen /goal job"]
w1 --> w2 --> w3 --> w4
end
OUT ~~~ EVT
EVT ~~~ WAKE
classDef hot stroke:#8f7ac9
```
---
# Sharing a container
Invite existing ZeroLog users with explicit read-only or read-write access.
## Current availability
Owners on plans with sharing (standard configuration: Pro, Max and Ultra) choose who can access a container. Read-only terminal access cannot type; read-write access permits edits and commands. Treat that as code-execution trust.
## Keep personal credentials private
Invite someone by username, or create an invite link. The person must already have a ZeroLog account. A link carries a fixed role and expires after 7 days unless you choose otherwise; you can also limit how many people may use it. Whoever opens a link sees the owner, the container and the role before accepting.
Change the role, remove a member or revoke a link whenever you like. A member you removed cannot rejoin through an old link; only you can invite them again. Knowing a container ID grants nothing.
Personal native Claude Code and Codex credentials require a workspace owned by you that nobody else can write to. Keep shared work and personal vendor-account sessions in separate workspaces. When you invite someone, any Claude Code or Codex agent container of yours that has this workspace is stopped, so your personal logins are no longer mounted next to the new member.
See [containers](https://zerolog.ch/docs/containers), [native account setup](https://zerolog.ch/docs/byoagent) and [current plans](https://zerolog.ch/pricing). These are the implemented access rules, not a claim of a new two-account invitation test.
## How a container is shared
Mermaid source of the diagram on this page:
```mermaid
%% Sharing: how access is granted and limited. Facts from server/multiplayer.js, routes/containers.js, tiers.js, agent.js.
flowchart TB
o["Owner on Pro, Max or Ultra
opens the container's
sharing panel"]
subgraph WAYS["Two ways to invite"]
u["Invite by username.
Must be an existing
ZeroLog user"]
l["Create an invite link.
Role fixed, expires after
7 days by default,
optional use limit"]
c["Person opens the link,
sees owner, container
and role, then accepts"]
l --> c
end
subgraph ROLES["Role you grant"]
ro["Read-only: view files,
attach to the terminal,
cannot type"]
rw["Read-write: edit files,
run commands.
Code-execution trust"]
end
lock["Others can write here:
your personal Claude Code and
Codex logins are not mounted"]:::hot
ctl["You can change a role,
remove a member or revoke a link.
A removed member cannot
rejoin with an old link"]
o --> u
o --> l
u --> ro
u --> rw
c --> ro
c --> rw
rw --> lock
ro --> ctl
rw --> ctl
classDef hot stroke:#8f7ac9
```
---
# Android app
ZeroLog for Android 2.0: the chat you know, and an agent that works on your phone.
Version 2.0 is a native app. Chat, chat list, Compute, jobs and your container's files are drawn by the app itself and talk to zerolog.ch directly, not through a web page inside the app. You sign in once; the session key stays in Android's encrypted storage. The agent can run commands, work with files and control other apps after you allow Android's accessibility service. Each action asks first unless you turn approvals off.
## What the app does
### Four tabs
Chats, Aufträge (jobs, filtered by All, Phone and Server), Dateien (files) and Gerät (device). Compute, the phone agent, tools and settings sit under Gerät.
### Phone jobs
The agent shows a plan in the chat, you start it, and you can pause, stop or take over while it works. Details are in the section below.
### Agent on the phone
Runs Python 3.11 on the device, plus a shell and app control.
### Folder sync
Choose a folder to sync with your container's `/work` folder. Files over 400 KB and binary files are left out rather than cut short.
### Files app
Your container appears as a storage location, so any "choose a file" dialog can reach it.
### /goal in the agent tab
[/goal](https://zerolog.ch/docs/goal) works here too. The check runs on the server, the same as in the web chat.
### Your own CLIs
Claude Code and Codex run in Termux (from F-Droid or GitHub, not the Play Store) inside a small Ubuntu the app sets up, about 1 GB. Settings → CLI sign-in shows the setup line. Each tool signs in through its own page in Chrome; the app never sees the code.
### Signing out
Removes the sign-in token from the phone's secure storage. Version 2.0.0-a3 fixes a bug that left it behind.
### Crash reports
After a crash the app sends the error type and the line in its own code on the next start. No account, chat or device ID.
**Download and installation**
Needs Android 8.0 or later. The download needs a signed-in ZeroLog account; if needed, you are sent to sign in and returned to the app page. The current build is 2.0.0-a7, a pre-release. It installs over earlier 1.x versions and keeps your settings. It is signed with a new key: if you installed 2.0.0-a1, uninstall that once first. Later updates install over a3. The [Android app page](https://zerolog.ch/mobile) lists the file, its SHA-256 checksum and the checksum of the signing key; the previous version, 1.21.7, stays available there.
[Open the Android app page](https://zerolog.ch/mobile)
## Phone jobs
Ask the agent in a chat for something that needs your phone. It answers with a plan card. Nothing runs yet.
1. Tap **Auftrag vorbereiten** (prepare job). The app lists the apps the plan needs. They form the job's allow-list, and apps the app protects are never on it. You can remove an app, add one with **App auswählen**, say whether the agent may ask you questions, and choose whether screenshots go to the AI (**Bilder an KI**).
2. Start the screen share. Android asks for your consent. From Android 14 it offers only the whole screen.
3. While the job runs, a small overlay shows the steps with **Pausieren** (pause), **Stoppen** (stop) and **Übernehmen** (take over). Pause holds at the next safe point, between two steps, and ends the screen share; continuing asks Android for consent again. Stop ends the job. Take over ends the agent's control and leaves the phone to you.
4. When the job ends, a result screen shows what happened. **Im Workspace weiterarbeiten** opens the chat that holds the plan and the steps.
The agent reads or taps only in apps on the allow-list. Anything else in the foreground is blocked, even with Privacy Guard switched off. During a job it cannot use the shell, send intents, open the browser or touch files. Screenshots leave the phone only when **Bilder an KI** is on, the screen share is active and an allowed app is in front. They go to the model; a model that cannot read images gets a note instead.
Limits: on some system screens, such as Settings and Files, Android hides overlays from other apps, so the panel is not visible there. Pause and stop then work from the app, and ending the screen share from the status bar pauses the job. A pause does not interrupt a step that is already running. How reliably each model proposes a plan has not been measured yet.
## Phone control
The agent can read the screen, tap, type and scroll in other apps once you allow Android's accessibility service. On Android 13 and later that switch is greyed out for apps from outside a store: tap it once, then App info → ⋮ → Allow restricted settings.
### Set up Claude Code and Codex in Termux
1. Install Termux from [F-Droid](https://f-droid.org/packages/com.termux/) or [the Termux GitHub releases](https://github.com/termux/termux-app/releases). The Play Store build cannot accept the app’s command handoff.
2. In the Android app, open Settings → CLI sign-in. Review and copy its setup line into Termux; it allows the app to start commands there.
3. Start setup. It prepares a small Ubuntu and supported CLIs. Allow about 1 GB and leave Termux running until setup finishes.
4. Choose sign-in beside a CLI. Follow that tool’s page in Chrome; the tool receives its own callback. Its account, limits and conversation remain separate from ZeroLog chat.
A connected tool appears in the agent’s model list and can use approved phone tools through the local bridge. Tools depend on the mode: the restricted phone-job mode above does not permit shell, intents or file access.
### Migration and file identity
The current pre-release package is `ch.zerolog.zerocode.debug`. The old web-only ZeroLog app is separate: keep it until sign-in to the native app works. For 2.0.0-a1, follow the signing-key migration on [Mobile’s file section](https://zerolog.ch/mobile#file). Keep an export of important data before uninstalling an app. Updates from earlier 1.x versions preserve settings as described on the download page.
### If setup fails
Write to [info@zerolog.ch](mailto:info@zerolog.ch) with your Android version, the step and visible error. For CLI setup, include only relevant last log lines after removing credentials, private paths and project content. Current and previous APK links and checksums are on [Mobile](https://zerolog.ch/mobile#file). This is not a report of a new device test.
## A phone job in the Android app
Mermaid source of the diagram on this page:
```mermaid
%% Android app, phone job. Facts from the app page and /mobile (2.0.0-a3, build 2102). The 2.0 native source was not available for this check.
flowchart TB
ask["1. You ask the agent in a chat
for something on your phone"]
plan["2. Plan card in the chat
Nothing runs yet"]
prep["3. Prepare job: review the apps
(allow-list), questions,
screenshots to the AI on or off"]
share["4. Start the screen share
Android asks for consent"]
subgraph RUN["Job running"]
run["Agent reads and taps only
in allow-listed apps"]:::hot
ov["Overlay with the steps"]
end
pause["Pause: holds at the next safe
point, ends the screen share"]
stop["Stop: ends the job"]
take["Take over: the phone
is yours again"]
res["5. Result screen
Continue in the workspace"]
ask --> plan --> prep --> share --> run
run --- ov
ov --> pause
ov --> stop
ov --> take
pause -->|"continue asks
for consent again"| run
run --> res
stop --> res
take --> res
classDef hot stroke:#8f7ac9
```
---
# Claude Code and other CLIs
Use supported command-line tools with a vendor account or with models available through ZeroLog.
- Claude Code
- Codex
Available tools are Claude Code and Codex. Their sign-in and usage follow the mode you choose.
## Native accounts and catalogue tools
Choose the connection mode first. Native Claude Code and Codex use vendor accounts; ZeroLog Code uses catalogue models.
*Setup, billing and limits for each mode.*
| | Your own vendor account | Models on this service |
| --- | --- | --- |
| How you set it | Use the native Claude Code /login or Codex device sign-in in a private workspace. | Use ZeroLog chat or ZeroLog Code and select a model from the catalogue. |
| Who bills you | The vendor, under your subscription. | Your ZeroLog credit, per request. |
| Which limits apply | That vendor's plan limits, unchanged. This service does not raise them, extend them, or send a request past them. | Your ZeroLog balance and account limits. |
| If a request fails | The vendor's limits apply. If a request is refused, follow the vendor's message. | If a request fails, choose another available model and try again. |
Choose the account you want to use
Your vendor subscription and ZeroLog usage have separate billing and limits. A ZeroLog account limit does not change the limits set by your vendor.
Models on this service
## Use a ZeroLog model
Choose a model from the catalogue. Requests count toward your ZeroLog usage and account limits.
With ZeroLog models, Claude Code runs through a translator that converts its Anthropic-style requests for the catalogue model you chose. Anthropic does not officially support pointing Claude Code at non-Claude models through a gateway, so this path can stop working after a Claude Code update. If it does, run the tool on your own vendor account instead. Anthropic models are not part of the ZeroLog catalogue; to use them, sign in with your own Claude subscription or API key.
## Where usage appears
*Your own vendor account*
Your vendor tracks this usage under its account and subscription.
*Models on this service*
ZeroLog shows this usage with your other account activity. Your account limits apply.
The [/goal](https://zerolog.ch/docs/goal) continuation command belongs to ZeroLog chat; it is not automatically a native Claude Code or Codex command. See [native sign-in](https://zerolog.ch/docs/byoagent) and [tool handover](https://zerolog.ch/bring-your-own-ai-subscription). A native login alone does not establish access to every vendor model. Settings, then AI accounts, shows whether each vendor account is connected and starts the Codex sign-in. Claude Code always signs in inside the CLI.
## Two ways to connect a CLI
Mermaid source of the diagram on this page:
```mermaid
%% CLI modes: two ways to connect, each with its own bill and limits. Facts from server/agent.js, routes/agent.js and the byoagent page.
flowchart TB
start{"Pick the connection
mode first"}:::hot
subgraph V["Your own vendor account"]
v1["Private workspace you own.
Claude Code: /login in the CLI.
Codex: device sign-in"]
v2["Vendor bills you.
Vendor limits apply,
unchanged"]
v3["Request refused:
follow the vendor's message"]
v1 --> v2 --> v3
end
subgraph Z["Models on ZeroLog"]
z1["ZeroLog chat or
ZeroLog Code.
Pick a catalogue model"]
z2["Your ZeroLog credit pays,
per request. Your account
limits apply"]
z3["Request fails:
pick another model"]
z1 --> z2 --> z3
end
start -->|native CLI| v1
start -->|catalogue model| z1
classDef hot stroke:#8f7ac9
```
---
# API reference
Call ZeroLog models from your own code with an OpenAI-compatible endpoint, and start work in your containers from other services with webhooks. 8 endpoints, all listed on this page.
- **Base URL**: `https://zerolog.ch`
- **Specification**: [/openapi.json](https://zerolog.ch/openapi.json) · OpenAPI 3.1.0
- **Version**: 1.1.0
- **Format**: JSON, UTF-8. Streams as server-sent events.
## Quickstart
You need a ZeroLog account. The token belongs to that account; every request you make with it is billed to your balance.
1. **Get your token.**
Sign in at zerolog.ch, then open [`/api/cli/token`](https://zerolog.ch/api/cli/token) in the same browser. Copy the value of `token`. If you use ZeroLog Code, `zerolog login` fetches it for you.
Your token, in the shell
```
export ZEROLOG_TOKEN="zl_YOUR_TOKEN"
```
2. **Check that it works.**
A `200` with a model list means the token is valid. A `401` means it is not.
List models
```
curl -sS https://zerolog.ch/api/cli/v1/models \
-H "Authorization: Bearer $ZEROLOG_TOKEN"
```
3. **Send a request.**
Any model your account may use works here. Without `model`, the catalogue default is used.
Ask a model
```
curl -sS https://zerolog.ch/api/cli/v1/chat/completions \
-H "Authorization: Bearer $ZEROLOG_TOKEN" \
-H "Content-Type: application/json" \
-d '{"model":"featherless/google/gemma-4-31b-it","messages":[{"role":"user","content":"Write a haiku about Swiss trains."}]}'
```
The endpoint speaks the OpenAI wire format, so OpenAI client libraries work if you set the base URL:
Python, OpenAI SDK
```
import os
from openai import OpenAI
client = OpenAI(
base_url="https://zerolog.ch/api/cli/v1",
api_key=os.environ["ZEROLOG_TOKEN"],
)
reply = client.chat.completions.create(
model="featherless/google/gemma-4-31b-it",
messages=[{"role": "user", "content": "Write a haiku about Swiss trains."}],
)
print(reply.choices[0].message.content)
```
## Authentication
Model endpoints take your client token in one of two headers:
- `Authorization: Bearer mhc-…`, which most clients send;
- `x-api-key: mhc-…`, for clients that send keys this way. If both are present, `x-api-key` wins.
The token has the form `mhc--.`. Each sign-in gets its own: the same browser or device always gets the same token, another one gets a different token. Tokens do not expire and stop working when the account is disabled. Treat a token like a password: anyone holding it can spend your balance.
To see and revoke your tokens, use **Settings → ZeroLog Code → Client tokens** in the app, or the three token endpoints below (`GET /api/cli/tokens`, `DELETE /api/cli/tokens/{id}`, `DELETE /api/cli/tokens`). The list shows the first characters of each token, never the full value. A revoked token gets `401` from the next request on; a response that is already streaming finishes. Revoking a paired device also revokes the tokens it fetched. Signing out does not revoke a token. Accounts that existed before 4 November 2026 may still have one older account token (`mhc--`); it shows up as `legacy` in the list and can be revoked the same way.
Webhooks carry their own token in the URL (`mhw_…` or `mhin_…`). It is shown once when you create the webhook, stored only as a hash, and revoked by deleting the webhook. See [Webhooks](https://zerolog.ch/docs/webhooks).
## Errors
Model endpoints answer errors in the OpenAI form: `{"error": {"message": "…", "type": "…", "code": "…"}}`. `code` is set when there is a machine-readable reason. A missing or wrong token gets `{"type": "error", "error": {"type": "authentication_error", …}}`. Webhooks answer `{"error": "…", "code": "…"}`.
*Error statuses of the model endpoints*
| Status | Type | What happened | What to do |
| --- | --- | --- | --- |
| `400` | `invalid_request_error` | The request cannot be served as sent: no model and no default, or the provider rejected it. | Fix the request. Do not retry unchanged. |
| `401` | `authentication_error` | Missing, malformed or wrong token, or the account is disabled. | Check the token. Do not retry. |
| `402` | `billing_error` | Balance used up, or the shared pool is empty. | Top up in Billing, then retry. |
| `403` | `permission_error` | Your account may not use this model (`model_not_available`), or no usable key exists for it (`no_key`; for `byok/…` models `byok_key_invalid`). | Pick another model, or fix your own provider key. |
| `404` | `not_found_error` | The model provider does not know the model. `upstream_status` carries its status. | Pick another model. |
| `408 · 409 · 413 · 422` | `invalid_request_error` | The provider answered with this client error, passed on unchanged. A `413` can also come from ZeroLog itself (body over 32 MB). | Read `error.message`; change the request. |
| `429` | `rate_limit_error` | The provider is rate-limiting. ZeroLog passes on `Retry-After` when the provider sent one. | Wait, then retry with backoff. |
| `503` | `api_error` | No provider could answer, inference on this provider is paused for customers (`provider_temporarily_paused`), or the provider has no working key on our side (`provider_unavailable`). | Retry later or pick another model. |
A provider's `401` or `403` concerns ZeroLog's key, not yours, so you get `503` instead and your client keeps its token. When a provider answered, its status is in `error.upstream_status` and in the `X-ZeroLog-Upstream-Status` header.
**Fallback instead of failure.** If the model you asked for fails, ZeroLog may answer with another one. You can tell from the `X-ZeroLog-Served-Model`, `X-ZeroLog-Rerouted-From` and `X-ZeroLog-Model-Notice` headers, and from a one-line notice at the start of the answer. Send `X-ZeroLog-Reroute: off` when you would rather get the error.
## Rate limits and sizes
- **Model endpoints** have no fixed request rate. What limits you is your balance (`402`) and the provider's own rate limits (`429` with `Retry-After` when the provider sends it). Request bodies may be up to 32 MB.
- **Goal webhooks**: 1 to 60 triggers per minute (default 6) and 1 to 1000 per day (default 100), set per webhook. Bodies up to 32 kB.
- **Inject webhooks**: 1 to 120 events per minute (default 60), set per webhook. Bodies up to 256 kB.
Webhook limits answer `429` with `retry_after_s` in the body. Back off for that long.
## Pagination
None of the public endpoints page their results. `GET /api/cli/v1/models` returns its whole list in one response.
## Streaming
Send `"stream": true` to receive the answer as `text/event-stream`. Each `data:` line is one chunk in OpenAI form; the stream ends with `data: [DONE]`. ZeroLog asks the provider for a final chunk with `usage`. If you close the connection early, ZeroLog stops the provider request and bills what was produced up to then.
Streamed answer (shortened)
```
data: {"id":"chatcmpl-example-2","object":"chat.completion.chunk","model":"featherless/google/gemma-4-31b-it","choices":[{"index":0,"delta":{"content":"Lake Geneva"},"finish_reason":null}]}
data: {"id":"chatcmpl-example-2","object":"chat.completion.chunk","choices":[],"usage":{"prompt_tokens":21,"completion_tokens":9,"total_tokens":30}}
data: [DONE]
```
## Webhooks and WebSockets
ZeroLog does not call your servers. The two webhooks are inbound: you give their URL to GitHub, Stripe, Zapier or your own service, and they call ZeroLog. A goal webhook starts the goal run it was created for. An inject webhook writes each event as a JSON line into `/work/.webhooks/incoming.jsonl` in its container, where an agent can read it. Both accept only `POST`; any other method gets `405`, so a link opened in a browser starts nothing. Before an event is stored, `Authorization`, `Cookie` and other credential headers are replaced by `‹redacted›`.
There are no public WebSocket endpoints. Terminals, the browser extension and the desktop link use WebSockets inside the signed-in app; they are not part of this API.
## Endpoints
Every endpoint below is also in [/openapi.json](https://zerolog.ch/openapi.json), so you can load it into Postman, Insomnia or a code generator. The examples use invented data.
### Models
OpenAI-compatible inference. Billed to your ZeroLog balance, or to your own provider key for `byok/…` models.
### GET `/api/cli/v1/models`
List models
OpenAI-shaped model list for clients that probe `/models`. Without `details` it returns the model your account resolves to by default, if your account may use it. The list can be empty; you can still send any model id from the catalogue that your account may use.
With `details=1` it returns every model your account can find in the web model picker (your own BYOK models, the recommended models and every usable model of the full catalogue), page by page, with name, family, context length, capabilities and price. Follow `next_cursor` until it is `null`. If the list is rebuilt while you page, the next page answers 409; start again without `cursor`.
- **Authentication**: Client token: `Authorization: Bearer mhc-…` or `x-api-key`
- **Parameters**: - `details` query — `1` returns the full list with metadata (paged). - `limit` query — Page size with `details=1`. Default 2000, maximum 5000. - `cursor` query — `next_cursor` of the previous page.
*Responses*
| Status | Meaning |
| --- | --- |
| `200` | Model list. |
| `400` | Invalid cursor. |
| `401` | Missing, malformed or wrong client token, or the account is disabled. A request without a token in the expected form and with a body over 64 kB is refused before it is read, with `{"error": "unauthorized"}`. |
| `409` | The model list changed while paging. Start again without `cursor`. |
| `503` | Model catalogue temporarily unavailable. |
curl
```
curl -sS "https://zerolog.ch/api/cli/v1/models" \
-H "Authorization: Bearer $ZEROLOG_TOKEN"
```
JavaScript
```
// Node 18+ as an ES module (example.mjs), or a browser console
const res = await fetch(`https://zerolog.ch/api/cli/v1/models`, {
method: "GET",
headers: { Authorization: `Bearer ${process.env.ZEROLOG_TOKEN}` },
});
console.log(res.status, await res.json());
```
Python
```
import os, requests
res = requests.get(
f"https://zerolog.ch/api/cli/v1/models",
headers={"Authorization": f"Bearer {os.environ['ZEROLOG_TOKEN']}"},
)
print(res.status_code, res.json())
```
Response 200
```
{
"object": "list",
"data": [
{
"id": "featherless/google/gemma-4-31b-it",
"object": "model",
"created": 0,
"owned_by": "zerolog"
}
]
}
```
### POST `/api/cli/v1/chat/completions`
Create a chat completion
OpenAI-compatible chat completions. The request body is passed to the model provider in OpenAI form, so the usual fields (`messages`, `tools`, `tool_choice`, `temperature`, `max_tokens`, …) work. Only `model`, `messages` and `stream` are interpreted by ZeroLog itself.
**Model.** Without `model`, the operator's catalogue default is used. A model your account may not use is refused with `403 model_not_available` before any provider is called. `byok//` runs on your own provider key and never falls back to the shared pool.
**Fallback.** If the chosen model fails upstream, ZeroLog may answer with another model. It then sets the `X-ZeroLog-Served-Model`, `X-ZeroLog-Rerouted-From` and `X-ZeroLog-Model-Notice` headers and puts a one-line `[ZeroLog] … is unavailable` notice at the start of the answer. Send `X-ZeroLog-Reroute: off` to get the error instead.
**Streaming.** With `"stream": true` the answer is `text/event-stream` in OpenAI chunk format, ending with `data: [DONE]`. ZeroLog asks the provider for a final usage chunk.
**Billing.** Each answer is billed to your balance (or your own key for `byok/…`). A stream that the client aborts is still billed for what was produced.
- **Authentication**: Client token: `Authorization: Bearer mhc-…` or `x-api-key`
- **Parameters**: - `X-ZeroLog-Reroute` header — `off` disables the automatic fallback to another model. You then get the provider's error instead of a different model.
- **Request body**: OpenAI chat completions request. Maximum body size 32 MB. - **`model` string**: Model id. Optional; without it the catalogue default is used. - **`messages` array, required**: - **`stream` boolean**: - **`tools` array**: OpenAI tool definitions, passed to the provider. - **`tool_choice` any**: Passed to the provider. - **`temperature` number**: - **`max_tokens` integer**:
*Responses*
| Status | Meaning |
| --- | --- |
| `200` | The completion. JSON without `stream`, an event stream with `"stream": true`. |
| `400` | No model given and no default configured, the body is not valid JSON (then `{"error": "…"}` as a plain string), or the provider rejected the request (`upstream_status` is then set). |
| `401` | Missing, malformed or wrong client token, or the account is disabled. A request without a token in the expected form and with a body over 64 kB is refused before it is read, with `{"error": "unauthorized"}`. |
| `402` | Your balance is used up, or the shared pool is empty. |
| `403` | Your account may not use this model (`model_not_available`; an admin session without a second factor gets `admin_mfa_required` with a step-up `url` instead when the model is admin-only), or no usable key is available for it (for `byok/…`: your own key is missing or invalid). |
| `404` | The provider answered with a client error, passed on unchanged (404, 408, 409, 422 and other 4xx). `upstream_status` carries the provider's status. |
| `413` | The request body is larger than this endpoint accepts, or the provider refused it as too large (`upstream_status` is then set). |
| `429` | The provider is rate-limiting. Retry after the time in `Retry-After` when present. |
| `503` | No provider could answer (provider down, timed out or refused our key), or customer inference on this provider is paused (`provider_temporarily_paused`), or the provider behind the model has no working key on our side (`provider_unavailable`). 401 and 403 from a provider are reported as 503, because they concern ZeroLog's key and not yours. |
| `4XX` | The provider answered with a client error, passed on unchanged (404, 408, 409, 422 and other 4xx). `upstream_status` carries the provider's status. |
curl
```
curl -sS -X POST "https://zerolog.ch/api/cli/v1/chat/completions" \
-H "Authorization: Bearer $ZEROLOG_TOKEN" \
-H "Content-Type: application/json" \
-d '{"model":"featherless/google/gemma-4-31b-it","messages":[{"role":"user","content":"Write a haiku about Swiss trains."}]}'
```
JavaScript
```
// Node 18+ as an ES module (example.mjs), or a browser console
const res = await fetch(`https://zerolog.ch/api/cli/v1/chat/completions`, {
method: "POST",
headers: { Authorization: `Bearer ${process.env.ZEROLOG_TOKEN}`, "Content-Type": "application/json" },
body: JSON.stringify({
"model": "featherless/google/gemma-4-31b-it",
"messages": [
{
"role": "user",
"content": "Write a haiku about Swiss trains."
}
]
}),
});
console.log(res.status, await res.json());
```
Python
```
import os, requests
res = requests.post(
f"https://zerolog.ch/api/cli/v1/chat/completions",
headers={"Authorization": f"Bearer {os.environ['ZEROLOG_TOKEN']}"},
json={
"model": "featherless/google/gemma-4-31b-it",
"messages": [
{
"role": "user",
"content": "Write a haiku about Swiss trains."
}
]
},
)
print(res.status_code, res.json())
```
Response 200
```
{
"id": "chatcmpl-example-1",
"object": "chat.completion",
"created": 1793700000,
"model": "featherless/google/gemma-4-31b-it",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "Rails through the pass\nsnow on the windows, warm tea\nthe clock is on time"
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 16,
"completion_tokens": 19,
"total_tokens": 35
}
}
```
### Authentication
Getting, listing and revoking the client tokens for the endpoints above.
### GET `/api/cli/token`
Get your client token
Returns a client token for the current sign-in and the base URL to use it with. Needs a browser session on zerolog.ch (open the address while signed in) or a paired device; ZeroLog Code fetches it with `zerolog login`.
Each sign-in gets its own token: asking again from the same sign-in returns the same token, another browser or device gets a different one. Tokens do not expire. They stop working when you revoke them (see the token list below), when you revoke the device that fetched them, or when the account is disabled. Signing out does not revoke them; use the list for that. After a revocation, the same sign-in gets a new token here.
- **Authentication**: Browser session (signed in at zerolog.ch)
- **Parameters**: - `name` query — Optional name shown in your token list, up to 60 characters. Without it, the device name or `Browser` is used.
*Responses*
| Status | Meaning |
| --- | --- |
| `200` | Your token and base URL. |
| `401` | Not signed in, or the account is disabled. |
| `403` | The session may not do this. `code` says why: `session_required` for a service or container credential, `cross_site` when another site started the request (open the address directly on zerolog.ch), or another code for a session that may not act. |
| `429` | More than 30 token requests in a minute from this account (`rate_limited`). `retry_after` and the `Retry-After` header say how many seconds to wait. |
Open this address in a browser where you are signed in. Scripts cannot use it: they would need your session cookie.
Response 200
```
{
"token": "zl_YOUR_TOKEN",
"token_id": "n17",
"base_url": "https://zerolog.ch/api/cli/v1",
"note": "Bearer token for the zerocode terminal client. Pick the model in zerocode; it is sent as body.model."
}
```
### GET `/api/cli/tokens`
List your client tokens
Lists the client tokens of your account that still work, newest first. The answer never contains a token value, only its first characters. Only your own tokens are listed; there is no way to address another account. A paired device can call it with its device token as a Bearer instead of the cookie.
- **Authentication**: Browser session (signed in at zerolog.ch)
*Responses*
| Status | Meaning |
| --- | --- |
| `200` | Your tokens. |
| `401` | Not signed in, or the account is disabled. |
| `403` | Not a sign-in that can hold client tokens, for example a service or container credential (`session_required`). |
| `429` | More than 30 token requests in a minute from this account (`rate_limited`). `retry_after` and the `Retry-After` header say how many seconds to wait. |
Open this address in a browser where you are signed in. Scripts cannot use it: they would need your session cookie.
Response 200
```
{
"tokens": [
{
"id": "n17",
"name": "zerocode: laptop",
"prefix": "mhc-1042-n17.Xq3v…",
"created_at": 1793700000000,
"last_used_at": 1793786400000,
"current_session": false,
"legacy": false
},
{
"id": "n12",
"name": "Browser",
"prefix": "mhc-1042-n12.b7Rk…",
"created_at": 1793600000000,
"last_used_at": null,
"current_session": true,
"legacy": false
}
]
}
```
### DELETE `/api/cli/tokens`
Revoke all your client tokens
Revokes every client token of your account, including the legacy one. Takes effect on the next request: a response that is already streaming finishes, the request after it gets `401`. Your browser session stays signed in. A paired device can call it with its device token as a Bearer instead of the cookie.
- **Authentication**: Browser session (signed in at zerolog.ch)
*Responses*
| Status | Meaning |
| --- | --- |
| `200` | Revoked. |
| `401` | Not signed in, or the account is disabled. |
| `403` | Not a sign-in that can hold client tokens, for example a service or container credential (`session_required`). |
| `429` | More than 30 token requests in a minute from this account (`rate_limited`). `retry_after` and the `Retry-After` header say how many seconds to wait. |
Open this address in a browser where you are signed in. Scripts cannot use it: they would need your session cookie.
Response 200
```
{
"ok": true,
"revoked": 3
}
```
### DELETE `/api/cli/tokens/{id}`
Revoke one client token
Revokes one of your client tokens by its id from the list. Takes effect on the next request with that token; in-flight responses finish. Your other tokens keep working. An id that is unknown, already revoked or belongs to another account gets the same `404`. A paired device can call it with its device token as a Bearer instead of the cookie.
- **Authentication**: Browser session (signed in at zerolog.ch)
- **Parameters**: - `id` path, required — Token id from the list: `n` and a number, or `legacy`.
*Responses*
| Status | Meaning |
| --- | --- |
| `200` | Revoked. |
| `401` | Not signed in, or the account is disabled. |
| `403` | Not a sign-in that can hold client tokens, for example a service or container credential (`session_required`). |
| `404` | No such token for this account. |
| `429` | More than 30 token requests in a minute from this account (`rate_limited`). `retry_after` and the `Retry-After` header say how many seconds to wait. |
Open this address in a browser where you are signed in. Scripts cannot use it: they would need your session cookie.
Response 200
```
{
"ok": true,
"revoked": 1
}
```
### Webhooks
Inbound webhooks. You create them in the app (Automation and container settings); the token in the URL is the only credential and is shown once.
### POST `/api/hook/{token}`
Trigger a goal webhook
Starts the goal run that the webhook was created for, in its chat and container. Created in the app; the token (`mhw_…`) is shown once and stored only as a hash.
The call returns `202` as soon as the run is queued. `status_url` points to the job in the app and needs your browser session; it is not a public endpoint.
Every method other than POST answers `405`, so that opening the link in a browser never starts a paid run. Maximum body size 32 kB.
**Limits** are set per webhook: 1–60 triggers per minute (default 6) and 1–1000 per day (default 100). The account needs the goal mode permission and credit.
- **Authentication**: The token in the URL. No header.
- **Parameters**: - `token` path, required — Webhook token, `mhw_` followed by 32 URL-safe characters.
- **Request body**: JSON. - **`prompt` string**: Appended to the stored objective when the webhook is in append mode; ignored otherwise. Longer text is cut at 4000 characters.
*Responses*
| Status | Meaning |
| --- | --- |
| `202` | Run queued. |
| `400` | The webhook's chat and container are no longer attached to each other (`not_attached`), or the body is not valid JSON (then without `code`). |
| `402` | No prepaid credit left on the account. |
| `403` | The account is disabled or lacks the goal mode permission. |
| `404` | Unknown or revoked token (`no_hook`). A wrong token and a revoked one get the same answer, so tokens cannot be probed. `no_chat` and `no_container` mean the webhook still exists but its chat or container is gone. |
| `405` | Any method other than POST. |
| `409` | The container could not be resumed for the run. |
| `413` | Body larger than this webhook accepts (32 kB for goal webhooks, 256 kB for inject webhooks). |
| `429` | Per-minute (`rate_min`) or daily (`rate_day`) limit of this webhook reached. Wait `retry_after_s` seconds. |
| `500` | Unexpected server error. The body never carries internal details. |
curl
```
curl -sS -X POST "https://zerolog.ch/api/hook/$ZEROLOG_HOOK_TOKEN" \
-H "Content-Type: application/json" \
-d '{"prompt":"Deploy finished for commit 3f2a91c. Check the release notes."}'
```
JavaScript
```
// Node 18+ as an ES module (example.mjs), or a browser console
const res = await fetch(`https://zerolog.ch/api/hook/${process.env.ZEROLOG_HOOK_TOKEN}`, {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({
"prompt": "Deploy finished for commit 3f2a91c. Check the release notes."
}),
});
console.log(res.status, await res.json());
```
Python
```
import os, requests
res = requests.post(
f"https://zerolog.ch/api/hook/{os.environ['ZEROLOG_HOOK_TOKEN']}",
json={
"prompt": "Deploy finished for commit 3f2a91c. Check the release notes."
},
)
print(res.status_code, res.json())
```
Response 202
```
{
"job_id": 5821,
"status_url": "/api/jobs/5821"
}
```
### POST `/api/inject/{token}`
Deliver an event into a container
Appends the request as one JSON line to `/work/.webhooks/incoming.jsonl` in the container the webhook belongs to. Point GitHub, Stripe, Zapier or your own service at this URL; an agent in the container can then read the events. The token (`mhin_…`) is shown once when you create the webhook.
Stored per event: time, method, request headers and the parsed JSON body. `Authorization`, `Cookie`, `X-Api-Key`, `X-Auth-Token` and `Proxy-Authorization` are replaced by `‹redacted›` before storage.
`202` is returned for every accepted event. `ok: false` with `status: "error"` means the event was counted but could not be written, usually because the container is not running.
Every method other than POST answers `405`. Maximum body size 256 kB. **Limit** per webhook: 1–120 events per minute (default 60).
- **Authentication**: The token in the URL. No header.
- **Parameters**: - `token` path, required — Webhook token, `mhin_` followed by 32 URL-safe characters.
- **Request body**: JSON.
*Responses*
| Status | Meaning |
| --- | --- |
| `202` | Event accepted. |
| `400` | The webhook's chat and container are no longer attached to each other (`not_attached`), or the body is not valid JSON (then without `code`). |
| `403` | The account is disabled or no longer holds container access (`sandbox`). Nothing is written to the container and no hit is recorded. |
| `404` | Unknown or revoked token (`no_hook`). A wrong token and a revoked one get the same answer, so tokens cannot be probed. `no_chat` and `no_container` mean the webhook still exists but its chat or container is gone. |
| `405` | Any method other than POST. |
| `413` | Body larger than this webhook accepts (32 kB for goal webhooks, 256 kB for inject webhooks). |
| `429` | Per-minute (`rate_min`) or daily (`rate_day`) limit of this webhook reached. Wait `retry_after_s` seconds. |
| `500` | Unexpected server error. The body never carries internal details. |
curl
```
curl -sS -X POST "https://zerolog.ch/api/inject/$ZEROLOG_INJECT_TOKEN" \
-H "Content-Type: application/json" \
-d '{"ref":"refs/heads/main","after":"3f2a91c","repository":{"full_name":"example-org/example-repo"}}'
```
JavaScript
```
// Node 18+ as an ES module (example.mjs), or a browser console
const res = await fetch(`https://zerolog.ch/api/inject/${process.env.ZEROLOG_INJECT_TOKEN}`, {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({
"ref": "refs/heads/main",
"after": "3f2a91c",
"repository": {
"full_name": "example-org/example-repo"
}
}),
});
console.log(res.status, await res.json());
```
Python
```
import os, requests
res = requests.post(
f"https://zerolog.ch/api/inject/{os.environ['ZEROLOG_INJECT_TOKEN']}",
json={
"ref": "refs/heads/main",
"after": "3f2a91c",
"repository": {
"full_name": "example-org/example-repo"
}
},
)
print(res.status_code, res.json())
```
Response 202
```
{
"ok": true,
"status": "stored",
"path": "/work/.webhooks/incoming.jsonl"
}
```
## Trying requests
This page has no "Try it" button. A live test would need a real token, and every call is billed, so a button in a public page would either spend someone's balance or need a sandbox account that does not exist yet. Copy the examples instead and run them with your own token.
## What this API does not cover
The rest of `/api/` belongs to the ZeroLog web app. It needs a browser session, changes without notice and is not meant for scripts. If you need something from there as a stable endpoint, tell us at [info@zerolog.ch](mailto:info@zerolog.ch). The Monitoring API is a separate public API with its own specification; both are listed in `/.well-known/api-catalog`.
The specification is checked against the server code on every deploy: a public endpoint that changes without its documentation stops the deploy.
---
# ZeroLog Code
Use ZeroLog from a terminal to work with your account, workspace files and models.
ZeroLog Code is based on OpenCode and connects the terminal client to your ZeroLog account.
Install ZeroLog Code, then run `zerolog login` and follow the pairing steps in your browser. Once connected, you can work with files in your workspace and use supported models from the terminal.
Install via shell script on Linux and macOS:
```
curl -fsSL https://zerolog.ch/zc/install.sh | sh
```
Or on Windows PowerShell:
```
irm https://zerolog.ch/zc/install.ps1 | iex
```
## Anonymous error reports
Versions of ZeroLog Code that include anonymous error reporting send a report when the client crashes or encounters an error during sign-in, connection or workspace sync. Reports include technical details about the failure and leave out email addresses, IP addresses, account credentials, local home-folder paths and exception messages.
To turn reporting off, set any one of these environment variables before starting ZeroLog Code: `ZEROLOG_TELEMETRIE=0`, `ZEROCODE_TELEMETRIE=0` or `DO_NOT_TRACK=1`.
| Feature | How it works |
| --- | --- |
| Device pairing | Run `zerolog login` and complete the pairing steps in your browser. |
| Workspace synchronization | Sync local changes with your workspace. If a transfer is interrupted, check the sync result and run it again for files that did not finish. |
| Session history | The session history is written to a chat on zerolog.ch, so you can read it there. The agent loop itself runs on your machine; closing the terminal ends the run. |
| Settings | Under Settings, ZeroLog Code you can set the default model for new sessions, switch container sync on or off by default and pick the default target container. The same screen lists your connected devices and lets you revoke access for one. |
| Usage | Runs using ZeroLog models follow your account limits and appear with your other usage. |
ZeroLog Code also supports Android Termux. For the Android app with phone-control features, see the [Android app guide](https://zerolog.ch/docs/app).
## From install to a running session
Mermaid source of the diagram on this page:
```mermaid
%% ZeroLog Code: pairing, local run, what reaches your account. Facts from server/anthropic-shim.js (/cli/v1/*), auth.js pairing, settings/zerocode-panel.tsx.
flowchart TB
i["1. Install ZeroLog Code"]
l["2. zerolog login prints
a pairing link"]
ap["3. You approve the link in your
signed-in browser. The device
gets a token"]
r["4. Agent loop runs on your machine.
Local tools stay local"]:::hot
subgraph ZL["Through your ZeroLog account"]
m["Models: limits and
usage as usual"]
s["Workspace sync to a
container, if switched on"]
h["Session history
shows up as a chat"]
end
i --> l --> ap --> r
r --> m
r --> s
r --> h
classDef hot stroke:#8f7ac9
```
---
# ZeroLog Desktop
A Windows app for zerolog.ch with chats, code, jobs, containers and VMs in one window. Same account and same server state as the web.
**Version 0.4.2.** It includes passkeys, account deletion and UI test access in the app, and memory entries can be picked by meaning. The installer is on the [Desktop product page](https://zerolog.ch/desktop). The download needs no sign-in.
This page is about the app. It is not [Desktop control](https://zerolog.ch/docs/desktop), where the agent works on your own PC over Tailscale and SSH. Desktop control is a separate web feature.
## What is in the app
### Connect with one click
- Mount a container's files as a Windows drive. This needs Windows 11 24H2.
- Open SSH into a container from the app or in Windows Terminal.
- Bring a container into your Tailscale network.
- The file access overview shows the SMB address and rotates the password.
### Compute
- Containers and Windows VMs in one view.
- Change the size, take snapshots, share a container and send invitations, set ports and your own domains.
- Create a VM from the image catalogue and use its desktop inside the app. Remote Desktop through the Windows client still works.
### Chat
- Slash commands, instructions and temperature per chat, and export.
- Import and archive, folders and multi-select.
- Dictation and read-aloud use the speech built into Windows, so the audio stays on your PC.
- Chats and memory can be encrypted. The keys stay on your device. If you lose both the passphrase and the recovery code, those entries are gone.
- Selection by meaning is optional. When encrypted memory is unlocked, the memory tab offers it. If you agree, the app downloads a language model of about 34 MB from zerolog.ch once, checks it against a fixed checksum and runs it on your PC. Without it, an entry only goes along when it contains your whole message.
### Code
- A coding agent in a local folder or in a cloud container.
- An editor with tabs and a conflict check.
- Git with branches, push, pull and review block by block.
- Terminal tabs.
- If something goes wrong, choose "Triage an issue" in the command palette. The agent writes a report on this PC, and "Save triage report" stores it as a file you can send to ZeroLog support.
### Jobs, models, security and Studio
Watch a job live and continue it with an extra budget you confirm; schedules, automations, webhooks and recipes are there too. Models come with the catalogue, the routing table and the prompt optimiser. Security has approvals, the vault and security scans. You can choose UI test access set up in the app when ordering a domain security scan in the app. The order confirmation says when that access may change data. Studio has designs with versions and MP4 export, the site builder, the image gallery and legal research.
### Passkeys
- Open Security, then Key proxy, then Passkeys. There you set up a passkey, add another one, remove one or reset them.
- Key-service approvals and vault actions that need a passkey are confirmed from the app as well.
- The passkey step runs in a small separate window on zerolog.ch. It uses Windows Hello or a security key and needs Windows 10 version 1903 or newer. The page in that window is in English, even when the app is set to German.
- We checked on Windows that the window opens and reaches the passkey prompt. We have not yet tested the Windows Hello step ourselves on a device with Hello. Passkeys still work on the website as before.
### Deleting your account and UI test access
Both work from the app. The app shows a code and opens your browser. Check that the browser shows the same code, then sign in there again. Nothing is confirmed inside the app alone, and the confirmation has to finish in the browser on the same computer.
### The rest of your account
Every area of your account is a native screen, and search finds it. Alt+1 to Alt+8 are keyboard shortcuts. You can make your own themes, with a contrast check, and let the app start with Windows.
## Platforms and requirements
- **Windows**: Windows 10 version 1809 or newer, or Windows 11, 64-bit.
- **Not available**: macOS and Linux builds, a Microsoft Store version, and opening a container in VS Code over Remote-SSH.
- **Download**: One installer file, about 170 MB. Built with Electron.
- **Account**: An internet connection and a zerolog.ch account. Containers need a plan with containers, from Mini.
## Signing in
The app signs in through your browser with device pairing. You confirm the pairing in the browser where you are already logged in to zerolog.ch. No password is typed into the app. The device token stays in the protected store of Windows. To revoke a device, open the web settings and go to Apps & devices, Devices.
## Updates
The app checks for new versions and installs them silently. Installed 0.3.0, 0.4.0 and 0.4.1 apps update themselves to 0.4.2, and you stay signed in.
## Installing
The installer installs per user and needs no admin rights. It asks for the install folder, a desktop shortcut, start with Windows and whether to open the app at the end, in German or English. To uninstall, use Windows Settings, Apps. Your data stays unless you tick the box to remove it.
The installer is not code-signed yet, so Windows SmartScreen warns once. The warning reads "Der Computer wurde durch Windows geschützt" (English Windows: "Windows protected your PC"). Choose "Weitere Informationen" ("More info"), then "Trotzdem ausführen" ("Run anyway").
To check the file first: [latest.yml](https://zerolog.ch/desktop/updates/latest.yml) names the current installer and its `sha512` value in base64. In PowerShell, in the folder with the installer, run `[Convert]::ToBase64String([Security.Cryptography.SHA512]::Create().ComputeHash([IO.File]::ReadAllBytes((Get-ChildItem .\ZeroLog-Desktop-Setup-*.exe | Sort-Object LastWriteTime | Select-Object -Last 1).FullName)))`. It must print the same value.
## Signing in to ZeroLog Desktop
Mermaid source of the diagram on this page:
```mermaid
%% ZeroLog Desktop: device pairing and where each part runs. Pairing per server/routes/auth.js (/pair/start, /pair/approve, /pair/claim).
flowchart TB
subgraph APP["ZeroLog Desktop (Windows app)"]
a["1. Ask zerolog.ch
for a pairing code"]
b["2. Show the code,
open the system browser"]
d["4. Poll until approved,
receive the device token"]
e["5. Keep the token in the
Windows protected store"]
f["Chat and Code call the
API with that token"]
end
subgraph BR["System browser"]
c["3. You confirm the code
where you are already
signed in"]:::hot
w["Login confirmation, previews,
payment and web-only areas"]
end
s["zerolog.ch
same account and server state
as the web"]
a --> b --> c --> d --> e --> f
f --> s
f -.->|"areas not native yet"| w
classDef hot stroke:#8f7ac9
```