Claude Code and other CLIs
Agent work happens on this machine rather than your laptop, and output is rendered the way the CLI intends.
- @anthropic-ai/claude-code
- @openai/codex
- @google/gemini-cli
The vendor command-line tools, installed in every container from npm at pinned versions and not modified.
Two ways to run them
You pick one per account, and they do not mix.
In the fallback mode this service’s models run the tool through a translator. Anthropic does not support routing Claude Code to non-Claude models through a gateway, so that path can break with a release — when it does you get a clear message, and running the tool on your own account (the other mode) is the way through. Anthropic-branded models are never served to the hosted tool on our credits: for those, sign in with your own Claude subscription or API key.
| Your own vendor account | Models on this service | |
|---|---|---|
| How you set it | Settings, AI accounts, then start the sign-in. The CLI prints its own login URL, you open it and paste the code back. | The default. Nothing to sign in to. |
| Who bills you | The vendor, on your subscription. Nothing is charged here, and the usage line says "not billed here" rather than CHF 0.00. | 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 spend limits, the budget governor, and your plan's run length. |
| Automatic fallback | None. A request the vendor refuses ends the turn with an error. | Yes. A failing request is retried, then the same request goes to another provider on this service's own keys. |
About the last row
The last row is the one people assume runs the other way. Provider failover, the free-model rotation and the spend ceilings all live in this service's own model path. When you are signed in with your own vendor account, none of them are in the loop: the request goes from your container to that vendor, and when that vendor says no, the answer is no.
Models on this service
The translator
In the second mode the CLI is pointed at a translator inside your container instead. Three variables do that, and they are set in this mode only.
- ANTHROPIC_BASE_URL
- The translator inside your container. It converts the Messages API to whatever the chosen model actually speaks.
- ANTHROPIC_AUTH_TOKEN
- A token derived from your account. It buys inference and nothing else.
- ANTHROPIC_MODEL
- Any model from the catalogue, not only Anthropic ones. Change it and the same CLI runs on a different model.
The translator listens on the container network only, never on the public internet. There is no external HTTP API to call from your laptop: the CLI runs in the container, which is also what lets it edit the files and run the commands it is reasoning about.
Where the spend and the record land
A run here reaches none of that, because no money of ours is in that loop. It is recorded and shown, and it is not stopped.
Spend through the translator lands in the usual place: the same per-request accounting, the same spend limits, the same budget governor.
Either way, a /goal run started from a
CLI is reported back to the site and appears under Jobs with the rest.