cursor opens a cookbook for its agent sdk

github.com/cursor/cookbook

TL;DR: cursor published a repository with examples for @cursor/sdk — a typescript wrapper around their coding agent. The sdk runs the agent locally (reading files from disk) or in a cloud vm with a cloned repository, streams run events, and supports model selection, cancellation, mcp, and persistent conversation state across runs. The cookbook ships four examples: quickstart, a prototyping tool, a kanban for cloud agents, and a terminal launcher.

What is in the repository

• quickstart — a minimal node script: create a local agent, send a request, iterate over the event stream.
• prototyping tool — a web app for scaffolding projects and iterating in a cloud sandbox.
• kanban board — an interface for browsing cloud agents grouped by status or repository, plus creating new ones.
• coding agent cli — running agents from the terminal.

sdk in short

npm install @cursor/sdk

import { Agent } from "@cursor/sdk";

const agent = await Agent.create({
  apiKey: process.env.CURSOR_API_KEY!,
  model: { id: "composer-2" },
  local: { cwd: process.cwd() },
});

const run = await agent.send("Summarize what this repository does");

for await (const event of run.stream()) {
  console.log(event);
}

The CURSOR_API_KEY comes from the cursor dashboard. One interface for three runtimes: local node, cursor cloud, self-hosted cloud. A single run is the unit of work for one prompt; the agent keeps context between them. There is Cursor.models.list(), run cancellation that aborts in-flight tool calls, and mcp servers (via inline config, a file, or the cursor dashboard).

Context

The cookbook ships against the backdrop of cursor’s broader pivot to agents as a first-class primitive: in april 2026 they shipped cursor 3 with an agent-first interface — parallel agents across repositories and environments, durable canvases in the side panel. In march came cursor automations — agent triggers from a commit, a slack message, or a timer. The sdk is the logical next step: what a button does inside the editor can now be invoked from ci, scripts, and third-party apps.

The closest comparisons are claude agent sdk and openai’s assistants/agents api. Cursor’s distinction is the primary binding to code and repositories: the cloud runtime clones the repository into an isolated vm, the local one reads the working directory. This fits batch tasks over a codebase (refactors, lint auto-fixes, generating a pr from a ticket) where the agent needs a real workspace, not just textual context.