Cursor's Agent SDK is out: build, test and deploy agents as code
Cursor made Agent SDK generally available today, so teams can keep an agent's prompt, tools and evals in Git and run it on Cursor or their own servers.
You can now build a Cursor agent as a normal code project: a Markdown prompt, TypeScript tools and evals, all in Git. Run it locally with a playground, then deploy it to Cursor or your own servers.
What changed#
Cursor posted at 2:58 PM ET: "The Agent SDK is now available." Its changelog says it is "generally available today." The docs say it uses "the Cursor SDK harness that powers Bugbot, Security Review, Approval Bot, and Rollouts."
Per the changelog, agents wake on pull request, comment and CI events from GitHub, GitLab and Bitbucket, on Slack mentions and DMs, on webhooks, and on a cron schedule. Starter templates include a security reviewer, an owners-policy approver, a PR autofixer, and a Linear or Jira triager.
Why it matters#
Prompts and tools become files your team can review and test like code. The docs place it next to Grok Bot: pick Grok Bot when "you want to configure and talk to an agent without writing code," and Agent SDK when you want versioned code around model judgment.
How to use it#
- Install Node 22.13 or newer, then run
npm install --global @grok/agent-sdk. - Run
agent-sdk init ./my-bot, thenagent-sdk devin that folder. - Add evals, run
agent-sdk eval, thenagent-sdk deployfrom a pushed Git ref.
We did not install the CLI or sign in to a Cursor account on this box; steps come from the docs only.
The docs also offer a ready-made setup prompt for a coding agent. A starter request (ours, untested):
Read https://cursor.com/docs/agent-sdk.md, then scaffold an Agent SDK project for a bot that [does what you want] when [where people reach it]. Add one smoke eval and show me the file plan before you write anything.
Good to know#
- The changelog says
npx @cursor/agent-sdk, but that package returned 404 on npm when we checked. The docs use@grok/agent-sdkand say it "was@cursor/bdkbefore." - Evals call your real tools. The docs say to "guard any writes first."
- Managed hosting doesn't work under Privacy Mode (Legacy): "
deploy, a hosted session, and an eval run all refuse." - Teams with caps default to 20 deployments and 100 deploys a day, per Limits.
- We found no pricing for hosted agents in the docs.
What to watch#
Pricing for hosted agents, and whether the changelog's package name gets fixed.
Last checked: Oct 26, 2026, 4:22 PM to 4:28 PM ET (Cursor X post; changelog and docs agent-sdk pages; npm view @cursor/agent-sdk E404 and @grok/agent-sdk 0.3.0). What changed: first published Oct 26; no updates since.