Starting work¶
Start work on an issue¶
Open Issues, choose the project and open an issue. The Requirements box at the top tells you whether the agent has what it needs, usually the repository. If something is missing, Analyze can post the question on the issue for you.
The buttons depend on where the issue is:
| State | Main actions | In the ⋮ menu |
|---|---|---|
| New | Analyze, Run pipeline | Implement, Prepare |
| Analyzed | Implement | Run pipeline, Prepare, Analyze again |
| Coded | Live preview, Publish | Fix / improve, Code again, Analyze again, Start over |
| Failed | Retry | Fix with instructions, Code again, Analyze again |
| Published | Live preview, Revise pull request | Code again, Analyze again |
Most of the time you want Run pipeline: it analyzes the issue, creates the branch, writes the code and runs the build and tests in one go. Tick Publish when done to have it open the pull request as well.
The other actions are for when you want more control:
- Implement codes an issue that's already been analyzed.
- Fix / improve keeps the current code and changes it the way you describe.
- Code again keeps the analysis but starts the code over from a fresh copy of the branch.
- Retry continues a failed run from the step that failed, with the error passed to the model.
Choosing a model¶
Every coding action opens a short dialog. The defaults are fine for most issues; open Advanced if you also want to pick the model that analyzes the issue.
Under the model you'll see an estimate such as ~45k tokens · ~6 min, based on similar runs that finished recently. Hover over it to see the usual range. The tag next to it tells you whether the model is free or paid. The first few runs in a new project show No estimate yet.
Using your own AI key¶
You can use Claude, GPT, Gemini or any OpenRouter model for your jobs, paid by you directly at the provider.
- Open My connections → AI models with your own key.
- Pick the provider, paste your API key (the Get a key link opens the provider's page) and click Add key.
- Choose up to eight of the key's models and click Save.
Those models now appear in the model picker as model · your key. Only you can use your key. Projects set to Free models only won't use it. Set a spending limit at the provider, as you would for any API key.
Following a job¶
Every action runs in the background as a job, so you can close the page. Activity lists your jobs, and the icon in the top bar shows how many are running.
Open a job to watch its log. A typical pipeline goes through Reading issue, Analyzing, Readiness, Coding, Building, Safety checks, Security review, AI review and, if you asked for it, Committing.
You can cancel a job; it stops at the next step. If the model goes offline, the job shows Waiting for model and continues by itself when the model is back.