Cursor Agent SDK: an agent as a code folder, with one-command deploy
The SDK is generally available. Define an agent in a project, test it with evals and trigger it from GitHub, Slack, webhooks or a schedule.
Cursor has made its Agent SDK generally available. An agent is now a regular project folder: you can test it, keep it in git and deploy it to Cursor’s infrastructure or your own.
What’s new
The folder holds everything: instructions.md with the prompt, tools/ with typed actions, skills/ with procedures, mcp-connections/ for MCP servers and channels/ for how the bot communicates.
Agents can be triggered by:
- pull request, comment and CI events from GitHub, GitLab and Bitbucket;
- Slack mentions and DMs;
- webhooks and cron schedules.
Follow-ups in a PR or Slack thread resume the same session.
How to try it
npx @cursor/agent-sdk init ./my-bot
agent-sdk devThe second command runs the bot locally with a playground. Templates cover security review, PR autofixing, and Linear or Jira triage. To deploy, run agent-sdk deploy.
How to test it
agent-sdk eval sends a fixed message and checks which tools ran and the reply’s structure. agent-sdk github replay runs a real PR through the bot and can save it as test fixtures.
Limitations
Evals check actions and structure, not the quality of an agent’s conclusions, so review its first PRs by hand.
Prepared with AI from public sources.