Cursor has spent its life as an editor. A very good one. The Cursor SDK, now in public beta, changes what that means.

With it you get programmatic access to the same runtime, harness, and models running inside Cursor’s desktop app, CLI, and web client. The agent that used to live behind your keyboard now lives in your CI/CD pipeline, your backend service, or buried inside another product entirely. And that’s the real shift here. Coding agents stop being something a developer talks to. They become something a system calls.

Spin one up locally against your own machine, or send it to Cursor’s cloud, where it gets a dedicated VM, a cloned repo, and a configured dev environment. Cloud agents keep running after your laptop sleeps too, living up to the promise of agents that they will work when you are not. Reconnect later, stream the conversation from wherever it left off. When the work’s done, the agent opens a PR, pushes a branch, attaches a screenshot. You can also run it on self-hosted workers if code can’t leave your network, which for a lot of enterprise teams, it can’t.

What developers are actually buying access to is the agent harness, not just the model. With codebase indexing and semantic search, and MCP servers for plugging in external tools. Skills the agent picks up automatically from a repo folder. Hooks for watching and controlling the agent loop, and subagents that get spawned mid-task with their own prompts and their own models. Building that stack from scratch is real work, and most teams underestimate how much impact it can have.

Early customers are already putting this to work. The engineering team at Faire is running agents on the SDK to keep its codebase healthy without constant developer babysitting. Other teams are kicking agents off straight from CI to triage failed builds and patch PRs automatically. A few are going further by embedding Cursor directly into customer-facing products, so end users get an agent without ever touching Cursor itself.

Mitch Ashley at The Futurum Group frames this as Cursor stepping into a fight over the agent control plane. Expose runtime, sandboxing, and execution hooks programmatically, and you’re no longer just competing with other IDEs. You’re competing with CI/CD platforms, observability vendors, and cloud providers, all of whom want to own how agents actually run inside enterprise pipelines.

The SDK runs on standard token-based pricing and gives access to any model Cursor supports, including its own Composer 2, pitched as the cheap, fast default for most coding tasks. A public cookbook on GitHub ships four starter projects, from a bare-bones quickstart to a kanban board where dragging a card kicks off an agent that opens the pull request itself (a full task loop end-to-end, not just generating a code suggestion for a human to act on).