SDK Agent: Cursor defines an agent as a code folder, tested by evaluations and deployed in one command

Markdown Prompt, TypeScript Tools, Integrated Evaluations: Cursor's new SDK Agent treats the agent like any piece of the repository, edited in Git and deployable at Cursor or on its own servers.

Cursor has made available its Agent SDK, which makes an agent a simple directory of the project. There is a file instructions.md for the prompt, a folder tools/ for the typed actions that the model can call, a folder skills/ for procedures loaded on demand, MCP connections, and "channels" that describe where the agent receives its messages and where it responds. Everything can be read, tested, and versioned in Git like the rest of the code.

From local post to accommodation

The command agent-sdk dev opens a local sandbox, while agent-sdk deploy constructs the agent, assigns it a URL, and launches it on the Cursor framework; self-hosting is still possible. Agents can react to pull requests and events from CI on GitHub, GitLab, and Bitbucket, Slack mentions, any webhook, or a cron-like schedule.

Built-in evaluations send fixed messages to the agent and check which tools it triggers, and a command allows you to replay real pull requests in the form of signed webhooks to test without waiting for a new push. Four starting models are provided: security reviewer, approver according to code ownership rules, automatic PR corrector, and Linear or Jira ticket sorter.