Cursor Agent SDK explained: agents as a folder of code, no separate price, and how it differs from the Claude Agent SDK and OpenAI's Agents SDK

The Cursor Agent SDK is Cursor's framework for building agents as code: a folder of Markdown instructions and TypeScript tools that you can test, put in version control and deploy, either to Cursor's servers or your own. Cursor released it on October 20, 2026, and says it's generally available. It drew renewed attention on Hacker News on October 28. The name invites confusion, because Anthropic and OpenAI already ship products called an Agent SDK and an Agents SDK, and they do different jobs. For Anthropic's version, see our guide to building your first agent with the Claude Agent SDK; for OpenAI's hosted alternative, see our explainer on the Agents API.

This explainer is based on Cursor's October 20 launch post and changelog, its Agent SDK and Cursor SDK documentation, its pricing pages, the @grok/agent-sdk and @cursor/sdk packages on npm (inspected, not run), Anthropic's Agent SDK documentation and OpenAI's Agents SDK documentation, checked October 29, 2026.

The short version

• What it is: a framework for agents that live in your repository and respond to events: pull requests, Slack messages, webhooks or a schedule. Cursor says it builds its own Bugbot, Security Reviewer and Rollouts agents with it.
• How it installs: @grok/agent-sdk on npm, not the @cursor/agent-sdk shown in Cursor's blog post. Node 22.13 or newer.
• Where it runs: Cursor-managed hosting with one agent-sdk deploy command, or a single Node process on your own infrastructure.
• What it costs: no separate price. Cursor's SDK docs say runs use the same pricing and request pools as the IDE and Cloud Agents. Hosting charges aren't published.
• How it differs: the Claude Agent SDK embeds Claude Code's loop in an app you run; OpenAI's Agents SDK is an open-source library for orchestrating agents in your own code. Cursor's adds triggers, evals and hosting, tied to Cursor.

What is the Cursor Agent SDK?

Cursor's launch post, by Rustam Lalkaka, describes it as "the framework we use to build our own managed agents like Bugbot, Security Reviewer, and Rollouts." Its docs add Approval Bot to that list and say the SDK uses the same harness as the Cursor SDK. An Agent SDK project is a folder. Per Cursor's changelog and docs:

• instructions.md holds the agent's prompt;
• tools/ holds typed TypeScript functions the model can call;
• skills/ holds procedures the agent loads on demand;
• mcp-connections/ adds tools from MCP servers;
• channels/ defines how the agent receives and answers messages.

Channels are what make it an event-driven bot rather than a chat loop. Cursor lists pull request, comment and CI events from GitHub, GitLab and Bitbucket; Slack mentions and direct messages; webhooks from any provider; and a cron schedule. GitHub events "pull through your Cursor account, so no public webhook URL is required," the changelog says. Each conversation keeps its own session, so a follow-up comment on the same pull request resumes where the agent left off.

The command-line tool handles the rest: agent-sdk dev starts a local playground, eval runs evaluations, and github replay <pr-url> turns a real pull request into webhook payloads you can test against. Starter templates include a security reviewer and a Linear or Jira triager, and agents can hand work to subagents.

How do you install it, and why is it called @grok/agent-sdk?

Cursor's docs say to install @grok/agent-sdk, and that the package was previously named @cursor/bdk. The launch post's code sample imports from @cursor/agent-sdk, but no package by that name exists on npm as of October 29, 2026. The npm registry shows @grok/agent-sdk first published on October 22, two days after the launch post, with version 0.5.0 published on October 28. Its package page lists Cursor's GitHub repository as the source and Cursor's website as the homepage.

Cursor hasn't explained the @grok scope, but it fits Cursor's ownership. Cursor announced on August 14, 2026, that it "has officially been acquired by SpaceX," as our report on Cursor's September releases covers, and Grok is the model line from SpaceX's AI arm. Cursor's sources disagree on the default model. The configuration reference on Cursor's website says an agent's model "Defaults to grok-4.5 with effort=high and fast=true." The 0.5.0 npm package says otherwise: the reference docs bundled with it say the model "Defaults to grok-4.7 with reasoning_effort=high and fast=true," and its code sets the same grok-4.7 default. The launch post's example uses Cursor's composer-2.5 instead. Setting the model explicitly avoids the question.

The package README calls the SDK "the open-source version of the engine behind Bugbot, Security Review, Approval Bot, and Rollouts." Its license file says something narrower: "© Anysphere Inc. All rights reserved. Use is subject to Cursor's Terms of Service." Anysphere is the company behind Cursor. The source ships with the package, but the license isn't an open-source one.

Where do Cursor Agent SDK agents run?

In two places, according to Cursor's deployment docs:

• Cursor-managed hosting. agent-sdk deploy "builds the agent, gives it a URL, and starts it on our infrastructure," the launch post says. Deploys build a pushed Git ref, so uncommitted changes don't ship. Secrets are managed in Cursor's dashboard. Managed hosting requires Cursor's Privacy Mode; with the legacy Privacy Mode, deploys, hosted sessions and eval runs "all refuse," the docs say.
• Your own infrastructure. "Each agent runs as a single Node process that can be packaged in a container," per the launch post. Self-hosting needs a Cursor API key and a bearer token. The docs say to run it under Node 22.13 or newer, "never Bun."

Either way, the model turns go through Cursor. The quickstart says you need "a Cursor login for model turns," and the README says "Cursor runs the loop."

How much does the Cursor Agent SDK cost?

Cursor hasn't published a separate price. Its launch post, changelog, Agent SDK docs and pricing pages say nothing about Agent SDK charges, hosting fees or which plans can deploy agents.

The closest official statement is in the docs for the Cursor SDK, the lower-level TypeScript package the Agent SDK depends on: "SDK runs follow the same pricing, request pools, and Privacy Mode rules as runs from the IDE and Cloud Agents." Usage from a user's API key bills to that user's plan, and a service account's usage bills to the owning team. When Cursor launched the Cursor SDK in public beta on April 29, 2026, it said it "is billed based on standard, token-based consumption pricing." Whether hosted Agent SDK agents add a compute charge isn't stated.

Cursor's pricing page lists these plans:

Last verified: October 29, 2026

Cursor's pricing docs add that usage draws on a weekly pool shared "across everything on your subscription," that third-party models are billed at API list price from prepaid credit on Premium and above, and that Teams and Enterprise pay an extra $0.25 per million tokens on eligible third-party models. How an Agent SDK deployment draws on those pools isn't documented.

How does it compare with the Claude Agent SDK and OpenAI's Agents SDK?

They share a name but solve different problems. Anthropic's and OpenAI's are libraries for building an agent inside your own application. Cursor's is closer to a framework for team bots, with triggers, evals and optional hosting built in.

Last verified: October 29, 2026

A few details matter when choosing:

• Claude Agent SDK. It spawns a Claude Code subprocess with built-in tools, hooks, subagents, MCP, permissions, sessions, skills and plugins. Anthropic's hosting guide puts a minimal container at "roughly $0.05 per hour" and says token cost "typically dominates." Third-party products can't offer claude.ai login "unless previously approved."
• OpenAI Agents SDK. Its primitives are agents, handoffs or agents as tools, guardrails, sessions, tracing and sandbox agents. The Python docs say it defaults to gpt-5.6-luna when no model is set, that some tool features work "only with OpenAI Responses models," and that traces upload to OpenAI unless you change the processor.
• Cursor Agent SDK. Its strengths are the parts the other two leave to you: triggers, sessions per pull request or thread, replayable evals and a deploy command. The trade-off is lock-in; one Hacker News commenter wrote that "the idea of coupling an app to a particular vendor doesn't seem appealing at all" (community comment).

For a wider look at how Cursor, Claude Code and Codex compare as coding tools, see our comparison.

Who should use the Cursor Agent SDK?

Teams already paying for Cursor that want custom review, triage or approval bots, and want them versioned and tested like other code, are the obvious audience. Cursor's post says one team built a reviewer that checks pull requests against a lessons-learned knowledge base "in about a day," and that "The PGA Tour uses the Agent SDK to power their software development factory." (The post credits the accompanying quote to "PGA America.")

If you need an agent embedded in your own product, with models you choose and infrastructure you control, the Claude Agent SDK or OpenAI's Agents SDK fits better. If you want a vendor to host the loop for you outside Cursor, Anthropic's Managed Agents and OpenAI's Agents API are the comparable products.

What's confirmed, reported and uncertain

• Confirmed (Cursor): the October 20 launch and GA statement; the @grok/agent-sdk package name and the grok-4.5 default model in its online docs; the project layout, channels, CLI commands, templates and deploy options; Node 22.13+; the Privacy Mode requirement for hosting; the Cursor SDK billing statements; plan prices on October 29.
• Confirmed (npm, inspected not run): @grok/agent-sdk first published October 22, version 0.5.0 on October 28, its README and license text, and the grok-4.7 default in its bundled docs and code; no @cursor/agent-sdk package.
• Confirmed (Anthropic and OpenAI docs): each SDK's description, packages, hosting model, provider options and license.
• Uncertain: which default model Cursor intends, since its online docs and the 0.5.0 package differ; any hosting charge for Cursor-managed agents; which Cursor plans can deploy; how deployed agents draw on usage pools; whether the blog's @cursor/agent-sdk import will be corrected or published.