Cursor · Cloud Agents API · field brief
No. The key starts background coding jobs on GitHub. It does not replace the Cursor app. The pictures below map what we verified live on your seats.
Think hire a contractor, not open a chat window. You name a repo and a task. Cursor runs an agent in the cloud. GitHub gets a branch and can get a pull request.
We proved this end to end: key cli-cursor, model composer-2.5, smoke job on ve-agent-scratch, finished in about a minute, PR #4 opened. You still merge or close.
How to read it. The accent box is the API. Everything left of it is you (or your CLI). Everything right of it is GitHub. No repo on the left means nowhere safe to push. That is why ve-agent-scratch exists as a default pad.
People say “Cursor” for both the desktop app and this HTTP API. Mixing them is how you expect chat completions and get a PR bot instead.
Your machine showed two seats. The IDE login was a free Google account. The API key belonged to another email. Features do not cross automatically.
How to read it. If you need pair-programming on this laptop, use the IDE (or pi). If you need “go change that repo while I am elsewhere,” use the API. Asking the API for a chat model is the wrong door.
Only one band is owned by cli-cursor. The bands above and below need other logins, plans, or products.
Enterprise admin and analytics never answered on this key. Private-worker pending queues asked for a service account. The IDE band is a separate Google login on this machine.
cli-cursor pays for. GitHub access is required under it; IDE and Enterprise sit outside it.How to read it. If a feature lives on a grey band, buying or fixing the API key will not open it. Widen GitHub App access when repos are missing. Use the IDE login for in-app chat. Use Enterprise keys for admin and analytics.
Numbers and identities below came from live probes, not marketing pages.
cli-cursorustogether14422@…composer-2.5 onlycollectivewinca + 94 myblackbeanca)ve-agent-scratch under Collective Wincursor-agent launch → FINISHED + PRsubwaymusicianxyz@…myblackbeancaScratch pad: collectivewinca/ve-agent-scratch is the default repo for jobs with no product target. It holds design rules so agents can push without stranding work in a disposable VM.
Use this when someone claims “the API unlocks everything.” Match each row to the stack picture above.
| Feature | This API key | What that means in practice |
|---|---|---|
| Start a cloud agent on a repo | YES | POST /v0/agents or cursor-agent launch |
| List, status, stop, delete agents | YES | Full life cycle for jobs you started |
| List linked GitHub repos | YES | Only repos the Cursor GitHub App can see |
| Branch + PR with composer-2.5 | YES | Model runs inside the agent, not as open chat |
| TypeScript / Python agent SDK | YES | Same Cloud Agents capability, nicer client |
| Cursor IDE window | NO | Install and sign in to the app |
| In-app Chat, Tab, Composer UI | NO | App plan + that app’s login |
| Raw chat-completions HTTP API | NO | Docs: agents only, not OpenAI-style inference |
| Drop-in model for OpenCode / pi | NO | Wrong product surface |
| Admin, Analytics, Bugbot APIs | NO | Enterprise + different key scopes |
Key lives in the secrets bus. New shells load it. Then list, launch, poll.
source ~/.agentcookie/secrets/cursor/secrets.env
cursor-agent me
cursor-agent repos ve-work
cursor-agent launch --pr collectivewinca/REPO "task"
cursor-agent status bc-…
cursor-agent stop bc-…
Only a few agents can run at once on the current plan. Stop or delete finished runs before you start more.