/goal and /ultracode: agents that run without you

• You set the condition

  Up to 4,000 characters. “Make it better” never becomes true. Write something the checker can see in the chat, such as “the build prints no errors”.

• A separate model checks

  After every answer it reads the chat since the goal was set: up to 400 messages, limited to the last 16,000 characters. If the condition is not met, it says why, and the model is called again with that reason.

• It runs on the server

  Close the tab and the next round still starts. Reload the page and the lines are where you left them.

• Every round is a line in the chat

  After each round you see the checker’s reason. /goal on its own shows the current goal, its state and the rounds used. /goal clear ends it.

• Met

  The checker found the condition true in the transcript.

• Impossible

  The transcript shows a hard blocker the model cannot remove itself, such as an access right you must grant. Slow progress does not count.

• Stalled

  The checker named the same missing piece four rounds in a row. The run is going in circles.

• Capped

  25 rounds by default. The operator can change that number.

• Time limit

  The run reached the time limit of your plan (below). Your chat and files stay.

• Paused with an error

  The checker failed, or your credit is at zero. The goal stays set. Top up, then send “continue”.

One /goal run may continue without a stop for at most 4 hours on Mini, 8 hours on Pro and 12 hours on Max. Ultra has no limit. The limit applies per run, not per month: a new goal starts a new run. See plans for what each plan includes.

/ultracode is a switch, not a job. When it is on, the model in your chat may split a large task across several agents in your container when that is faster than doing it alone. It decides when. Short questions still get a short answer.

• Parts run in parallel

  The task is split into parts, and each part gets its own agent.

• Two checkers per part

  After an agent finishes, two reviewers are told to find what is wrong with the files it changed, not with its summary. A part they reject goes back for a fix.

• A card in the chat

  The plan, one line per agent with its state, and a log per agent. The card streams from the server, so you can reload or switch devices and it picks up where the run is.

• The same limits as any job

  Your plan’s run time and the budget governor apply. Several models on one task mean several bills, so use it where being wrong is expensive: audits, migrations, a review before a release.

• Credit is the hard stop

  Paid models stop at zero credit. Work does not continue and bill you later.

• A time limit per plan

  A /goal run ends at the time limit of your plan.

• A cap for delegated agents

  Delegated sub-agents draw from a default cap of CHF 0.50 per chat turn, with at most four delegated calls per turn.

• No spend cap per run

  There is no spend cap per run, per chat or per agent. A long run can use all the credit you have until it ends or the credit reaches zero.

• The budget governor is the operator’s

  It adds up what the whole service spends on models in a month. The operator sets it, not you, and it does not watch a single run.

Details: /goal, /ultracode, Spend limits, Budget governor, Jobs.

• How do I stop an autonomous AI agent from running up a big bill?

  Your credit is the hard stop: at zero, paid models stop, and nothing is billed afterwards. Each /goal run also has a plan time limit, 4 hours on Mini, 8 on Pro, 12 on Max and none on Ultra. There is no spend cap per run, so check your balance before a long run.

• What happens if a /goal condition is never met?

  The run stops on its own. It ends as stalled when the checker names the same missing piece four rounds in a row, as capped after 25 rounds by default, and as time limit when your plan’s run time is used up. Your chat and files stay, and you can set a new goal.

• Does a /goal run continue after I close the browser?

  Yes. The server drives the run, so the next round starts even with the tab closed. When you come back, reload the page: every round is a line in the chat with the checker’s reason. If your credit reached zero meanwhile, the goal is paused and waits for you.

• Is /ultracode more expensive than a normal chat?

  Usually, yes. When the model splits a task, several agents and two checkers per part each call a model, so one task produces several model bills. Use it where a wrong result is costly, such as audits or migrations. For everything else a single run is cheaper.

• Claude Code in the cloud

  Run long Claude Code tasks as /goal in your own container.

• AI coding agent with a persistent workspace

  Files, snapshots, jobs and schedules that outlast a chat.

• API key proxy for AI agents

  Let an agent use a key without ever seeing it.

• Use your own Claude or ChatGPT subscription

  Claude Code and Codex on the plan you already pay for.