Cursor puts its coding agents into an SDK

Coding agents are moving past the chat box on your laptop.

Cursor just released the public beta of its TypeScript SDK, letting developers programmatically spin up the exact same agent harness, runtime, and models that power Cursor's desktop, CLI, and web apps with just a few lines of code.

Cursor under the hood, now in npm

Available via npm install @cursor/sdk, the package exposes the full Cursor agent stack to programmatic workflows.

Developers can run agents locally, offload them to self-hosted workers inside private networks, or run them on Cursor's cloud infrastructure against isolated virtual machines.

In the cloud mode, agents keep working even if your machine disconnects. They can clone repositories, run tests in secure sandboxes, and automatically open pull requests with branches, screenshots, and demos attached.

Because the SDK connects directly to Cursor's Cloud Agents API, any run started from code automatically surfaces inside the Cursor desktop or web app, allowing human engineers to inspect progress or take over at any moment.

The SDK supports all frontier models available in Cursor, including its proprietary Composer 2.5 model and external models like GPT-5.5. It also imports Cursor's contextual engine, including codebase semantic search, instant grep, MCP server integrations, custom skills, and subagent delegation.

Early teams are already embedding these programmatic agents directly into CI/CD pipelines to debug failed builds, creating internal tools for non-technical teams, and integrating coding agents straight into their own software products.

Cursor also published several starter repositories, ranging from a basic Node.js quickstart to an agent-driven kanban board where dragging a card automatically triggers an agent to write the fix and open a PR.

Why it matters

Cursor is clearly pivoting from an AI-assisted code editor to the foundational operating layer for autonomous software engineering. Instead of waiting for a developer to ask a question in an IDE, software tasks can now be dispatched by events like a failed GitHub Action, a Slack alert, or a customer ticket.

We think this is a brilliant move that leapfrogs standard API wrappers. Setting up safe code sandboxes, durable background execution, and dynamic context indexing is brutal work that usually stops teams from building reliable custom agents. Cursor solved that messy plumbing for its own apps and is now renting it out via a token-based consumption model.

The real test will be cost and verification. Running automated cloud VMs and frontier models across continuous integration runs can rack up massive token bills fast. More importantly, when agents run without human eyes on every keystroke, teams will need bulletproof automated testing to ensure AI PRs do not quietly degrade codebases over time.