Skip to content
AI EmployeeSuper-AgentsAgent-to-AgentBenchmarksPricingBlog

Cursor Agent SDK: Agents as Code, Evals, Self-Hosting

At a glanceQuick answers
What is the Cursor Agent SDK?
A toolkit to define, test, evaluate and deploy agents as code: a folder of Markdown instructions and TypeScript tools that Cursor runs in its model loop, on Cursor-managed hosting or your own servers.
How is it different from Grok Bot or Cursor Automations?
Cursor’s own table: Automations are a prompt, model and trigger set in the dashboard; Grok Bot is configured in an app without code; the Agent SDK is a versioned Markdown and TypeScript project you review and test.
What does it cost?
Cursor’s docs do not list a separate price for the SDK or hosted deployments. Model turns need a Cursor login.
Editorial data illustration on a near-white ground: an open navy folder labelled AGENT AS CODE holding three cards, instructions.md, tools.ts and evals, with arrows out to a Slack bubble, a pull request icon and a server rack
Fig 0A folder you can version, test and deploy. Made by CellCog's image agent, running GPT Image 2.5.

Cursor made its Agent SDK available on October 26, 2026, a way to “define agents as code, with prompts in Markdown and tools in TypeScript.” Agents deploy on Cursor’s hosted infrastructure or run on your own servers, and built-in evals test every change.

The pitch is control. Where Grok Bot and Cursor Automations are set up in an app or a dashboard, an Agent SDK agent is a folder of files your team reviews, versions and tests like any other code.

On this page · 8 sectionsOpen
  1. What Cursor shipped
  2. Agent SDK, Grok Bot or Automations
  3. Evals are the point
  4. The package history
  5. What Cursor left out
  6. Our conflict, and who this is for
  7. What we are watching
  8. Sources
Key points5 · 5 min full read
  1. An open folder with a document and a code bracket.
    Cursor made its Agent SDK available on October 26, 2026: agents defined as code, with prompts in Markdown and tools in TypeScript, deployed on Cursor’s hosting or your own.
  2. Four small gears driven by one large gear.
    It runs on the same Cursor SDK harness that powers Bugbot, Security Review, Approval Bot and Rollouts, Cursor’s own agents.
  3. A clipboard with check marks beside a small robot.
    Evals are built in: each test sends a fixed input to the real agent and checks which tools ran and what the answer looks like.
  4. A package box with a new label over an old one.
    The package is @grok/agent-sdk on npm, renamed from @cursor/bdk, which first appeared on September 22; version 0.3.0 is marked GA.
  5. A price tag with a question mark.
    Cursor’s docs list no separate price for the SDK or for hosted deployments; only running a turn and running evals need a Cursor login.

§ 01What Cursor shipped

Cursor’s docs say the SDK “uses the Cursor SDK harness that powers Bugbot, Security Review, Approval Bot, and Rollouts”, the agents Cursor already runs for its own customers. You write the instructions and tools; Cursor runs the model loop that calls them.

  • Agents as code. A project keeps the agent under bot/: instructions.md is the always-on system prompt, each file in tools/ is one typed TypeScript tool, and skills, subagents, MCP connections, channels and hooks each get a folder.
  • Where people reach it. Cursor lists Slack, pull requests, issue trackers and Cursor itself.
  • Where it runs. Cursor-managed hosting with deploy, rollback, logs and per-deployment secrets, or your own infrastructure through the CLI’s serve command.
  • Who can touch it. Deployed agents carry three roles: Viewer opens the agent, Editor deploys, stops and restarts, Owner adds secrets and access.
  • Getting started. Node 22.13 or newer, npm install --global @grok/agent-sdk, agent-sdk login, then agent-sdk init and agent-sdk dev. Cursor also ships a prompt you paste into a coding agent to scaffold the first project for you.

§ 02Agent SDK, Grok Bot or Automations

Animated scene: a folder fills with three cards, instructions.md, tools.ts and evals; a run of green checks passes; then lines carry the agent out to Slack, a pull request and a server
Fig 1An agent is a folder: instructions, tools and evals, deployed where your team works

Cursor publishes its own comparison of the three ways to run an agent on its platform.

Cursor Automations Grok Bot Agent SDK
You define A prompt, model and trigger in the dashboard Instructions, skills and connected tools in the app A versioned Markdown and TypeScript project
Where it runs Cursor-managed Cursor-managed Cursor-managed hosting or your infrastructure
Choose it when The job fits a schedule or event trigger You want to configure and talk to an agent without code You want code around model judgment your team can review, test and improve
Table 1Cursor’s three agent products, from its Agent SDK docs (read October 26, 2026)

The third column is the news: Cursor now offers the same agent machinery behind Bugbot to any team willing to write the code.

§ 03Evals are the point

An eval sends a fixed input to the real agent and checks what it did. Cursor’s example passes only if the agent calls its inspection tool and does not approve a pull request. Evals can also check the shape of an answer against a schema, run in parallel, and get tagged so a quick smoke set runs on every change.

Cursor calls the loop “hillclimbing”: improve the agent, then lock each win with an eval so a later prompt change cannot quietly undo it. That is what separates this from a prompt pasted into a dashboard.

§ 04The package history

Date Event
September 22, 2026 First version published as @cursor/bdk
October 22, 2026 Version 0.2.27
October 23, 2026 Changelog marks 0.3.0 the GA release; 0.2.29 published
October 25, 2026 0.3.0 published to npm under @grok/agent-sdk
October 26, 2026 Cursor announces the Agent SDK on X
Table 2@grok/agent-sdk on npm (registry read October 26, 2026)

The @grok scope fits the company’s ownership: Cursor is sold inside The X Subscription alongside Grok and Grok Bot.

§ 05What Cursor left out

  • Price. The docs list no separate price for the SDK or for hosted deployments. Cursor’s plans page says paid plans share a weekly usage pool for Cursor models, and Premium at $30 a month adds Cloud Agents, Automations and Grok Bot; it does not yet name the Agent SDK.
  • License. The npm package points to its own LICENSE.md, not a standard open-source license.
  • Limits. Nothing on how many hosted agents a team can run or how hosted turns are metered.

§ 06Our conflict, and who this is for

Our conflict, declared: CellCog sells AI employees, a different answer to the same question of how a team puts agents to work, so treat our read with that in mind.

The Agent SDK is built for engineering teams that want an agent’s every instruction and tool in source control, with tests in front of each change, running in Slack and on pull requests. If that describes your team, it is the most serious offer Cursor has made to people building agents rather than using them. If you want an agent that holds a whole role without writing its code, that is a different product.

§ 07What we are watching

  • Pricing for hosted agents. How Cursor meters deployed turns and whether plans name the SDK.
  • The templates. Which review and triage agents teams publish on top of it.
  • Grok Bot overlap. Whether Grok Bot gains a path to export a configured agent into an Agent SDK project.

§ 08Sources

Frequently asked5 questions

Q1When did Cursor release the Agent SDK?

Cursor announced it on X on October 26, 2026. Its changelog marks version 0.3.0 as the GA release, dated October 23, and npm shows 0.3.0 published on October 25.

Q2How do I install the Cursor Agent SDK?

With Node 22.13 or newer, run npm install –global @grok/agent-sdk, then agent-sdk login, agent-sdk init to scaffold a project and agent-sdk dev to run it locally.

Q3Can I self-host Cursor Agent SDK agents?

Yes. Cursor says agents deploy on its hosted infrastructure or run on your own, and the CLI includes a serve command for running an agent server yourself.

Q4What can an Agent SDK agent connect to?

Cursor lists Slack, pull requests, issue trackers and Cursor itself, plus MCP connections, subagents, skills, hooks and secrets managed per deployment.

Q5Is the Cursor Agent SDK open source?

The npm package points to its own LICENSE.md rather than a standard open-source license, and Cursor does not describe it as open source.

Published 26 October 2026 All Choosing a platform →