Workflows
Cursor announces the Agent SDK
The official post: define agents in code, prompts in Markdown, tools in TypeScript.
Cursor’s account posted on 26 October 2026 that the Agent SDK was available. Agents are defined in code, prompts live in Markdown, tools are written in TypeScript, and you can deploy on Cursor’s hosting or run them yourself. Built-in evals are there so a change can be measured.
The post is an announcement, not the tutorial. Follow the docs it links, keep prompts and tools in version control, and keep secrets in the environment. Write evals from your own tasks, not from the demo in the announcement. Likes and reposts were captured on 15 November 2026.
This page is a guide, not the original text.
Note (verified 15 November 2026): This is a product-announcement post and may age quickly. This review did not reopen the original post, and it did not check whether Cursor has renamed the Agent SDK or changed its entry point.
What problem it solves
This catalog entry comes from a product announcement on the official Cursor account on 26 October 2026. It is not an open-source repository entry. The product to follow today is Cursor's Agent SDK: a way to define, test, evaluate, and deploy an agent as Markdown and TypeScript in version control. Prompts can live in bot/instructions.md, tools are TypeScript files under bot/tools/, and evals sit in evals/ at the project root. Official docs say it uses the same Cursor SDK harness that powers Bugbot, Security Review, and related products. The documentation is at cursor.com/docs/agent-sdk.
It is not the Agent or Chat panel inside the Cursor editor, and it is not the UI that edits the repository you currently have open. Official docs also contrast it with Cursor Automations and with Grok Bot, which you can configure without writing code. Agent SDK is the path to take when you want instructions, tools, and evals to be reviewable, versioned code, and when you may deploy to Cursor-managed hosting or to your own infrastructure.
Who it is for
It is for teams that already have a concrete agent job, such as reviewing pull requests or triaging issues, and that want prompts, tools, and regression evals in Git. You can develop locally with the official CLI, deploy with agent-sdk deploy onto Cursor-managed hosting, or self-host with agent-sdk serve as the docs describe. It is a poor fit if you only want to ask questions in a Cursor window and do not plan to maintain an agent project.
Finish login, scaffolding, and one run that has no external side effects before you connect GitHub, Slack, or other channels. Keep secrets and Cursor credentials in the environment or in the login state official docs describe. Do not commit them. This page follows the official docs and the npm package notice. It is not a reprint of the original announcement.
How to install it
Verified on 2026-11-16 (Asia/Shanghai). The commands come from the official docs. You need Node.js 22.13 or newer. The current package is undefined. The docs say it was previously named @cursor/bdk. On that day, npm listed both @grok/agent-sdk and @cursor/bdk at 0.6.0 (published around 00:45 Asia/Shanghai on 4 November 2026). Do not confuse it with another still-published package, undefined, whose registry description is "TypeScript SDK for Cursor agents." That is not the install package in the Agent SDK docs. @cursor/agent-sdk returned 404 on this review date.
npm install --global @grok/agent-sdk
agent-sdk login
agent-sdk init ./my-bot
cd my-bot
agent-sdk devagent-sdk login opens a browser sign-in. Official docs say local run and eval commands need a Cursor credential. The project is rooted at bot/, with evals in the adjacent evals/ directory. The command index is in the CLI reference. The older bdk command remains available.
The license is not an open-source SPDX identifier. npm lists SEE LICENSE IN LICENSE.md. The LICENSE.md inside the 0.6.0 tarball is titled "Cursor SDK License," credits Anysphere Inc., and says use is subject to Cursor's Terms of Service. This page only reports that package notice. It does not add a grant.
For self-hosting, official docs require agent-sdk serve on Node 22.13 or newer and say not to use Bun. Before a Cursor-managed deploy, the docs emphasize that deploy builds a pushed Git ref, so uncommitted changes do not ship. Those steps involve secrets and network exposure. Follow Deployment rather than treating this page as a production checklist.
A worked example
This example only scaffolds a local project and runs one introductory run, to confirm that the CLI and login work. Do not attach a real repository write in this step. Nothing was executed here, so this page does not invent the agent's reply.
- Confirm
node -vis 22.13 or newer, then runnpm install --global @grok/agent-sdk. - Run
agent-sdk loginand finish the browser sign-in. - Run
agent-sdk init ./pr-reviewer-demo && cd pr-reviewer-demo. - Open
bot/instructions.mdand replace it with a few lines of your own, for example: say that you are a demo reviewer, and that you must not touch external systems unless a tool call is explicitly requested. - In a second terminal, run:
agent-sdk validate --dir .
agent-sdk run --dir . --message "Introduce yourself in one sentence. Do not call any tools."- The expected effect is that
validatepasses the project checks, and thatrun, once you are logged in, returns a one-sentence introduction without calling a write tool. If login is incomplete, official docs say model turns fail. The fuller PR-reviewer walkthrough is in the Quickstart. Do not post a review on a real pull request from those GitHub examples until you have explicitly allowed that step.
Common pitfalls
- Treating Agent SDK as another name for the built-in Cursor editor Agent. Both use a model, but the SDK maintains a separate Markdown and TypeScript project, and the entry point is the
agent-sdkCLI, not the editor chat. - Installing the wrong package. Current docs require
@grok/agent-sdk(formerly@cursor/bdk).@cursor/sdkis a different package.@cursor/agent-sdkdid not exist on this review date. - Running
runorevalbefore you log in. Official docs say those commands need a Cursor credential. - Using a Node version below 22.13, or self-hosting with Bun. The runtime requirement is explicit.
- Running
deployafter local edits that were never pushed. Official docs say uncommitted changes do not ship. - Treating the proprietary license as an open-source license you can redistribute at will. The packaged LICENSE.md points at Cursor's Terms of Service.
FAQ
Is this the Cursor editor, or a separate agent SDK?
It is a separate Agent SDK for defining deployable agents as code. It can run on Cursor-managed hosting or on your own infrastructure, but it is not the Agent panel in the editor, and it is not a Cursor rules or skills entry this catalog might list elsewhere.
Why is the package named @grok/agent-sdk?
That is the current name on the official docs and on npm as of 16 November 2026. The docs say it was previously @cursor/bdk, and that older name still resolved to the same 0.6.0 release on this review date. This page follows the official docs and does not speculate about the rename.
May I redistribute the SDK freely?
Do not assume an open-source grant. npm points at the packaged LICENSE.md, which is titled Cursor SDK License and requires Cursor's Terms of Service. Read that file and the terms before you decide how far you can use it.
Is Grok Bot the same thing as Agent SDK in the official docs?
No. The official comparison table keeps them in separate columns. Grok Bot is for configuring instructions and tools in an app without writing code. Agent SDK is for a versioned Markdown and TypeScript project with review and evals. This page covers Agent SDK only.
Do I have to deploy onto Cursor-managed hosting?
No. You can start locally with agent-sdk dev and agent-sdk run. Official docs also offer agent-sdk serve for self-hosting, and agent-sdk deploy for managed hosting. Read the deployment docs for secrets, authentication, and the Node version before you self-host.
Sources and license
The catalog source for this entry is Cursor's official product announcement. The hands-on path is the Agent SDK documentation and the npm package @grok/agent-sdk. The license is the Cursor SDK License in the packaged LICENSE.md, and use is subject to Cursor's Terms of Service. This page is a paraphrase, for CQBase — Context & Query Base, of the official docs, the CLI reference, the Quickstart, and the npm metadata, checked on 16 November 2026. If the package name, runtime, or terms change, the official site wins.
In this category
- Workflows
Firecrawl
A web-crawling library for AI agents, licensed AGPL-3.0. The official MCP server is cataloged separately.
firecrawl - Workflows
MarkItDown
Microsoft’s Python tool for turning office documents into Markdown that a model or agent can read.
microsoft - Workflows
Claude Code
Anthropic’s terminal coding agent: it reads the repo, edits files, and runs commands.
anthropics