Cursor launches TypeScript agent SDK

Coding agents are starting to look less like chat features and more like infrastructure. That’s the real news in Cursor’s April 29 launch of a public-beta TypeScript SDK. Instead of only using Cursor inside the editor, developers can now call the same agent system from their own code, run it locally or in a cloud VM, and wire it into CI pipelines, automations, or product backends. (cursor.com) ### What actually launched? Cursor released `@cursor/sdk`, a TypeScript package that exposes the same runtime, harness, and models used in Cursor’s desktop app, CLI, and web app. The package is in public beta, and the basic pitch is simple: create an agent in code, send it work, stream events back, and manage the run without rebuilding the whole agent stack yourself. (cursor.com)K”? Because the hard part of coding agents usually isn’t prompting. It’s everything around the prompt — sandboxing, state, repo setup, reconnecting to long-running jobs, model swaps, and getting the thing to behave predictably in production. Cursor is basically saying: we already built that machinery for our own product, and now you can rent it instead of recreating it. (cursor.com) ### Where do these agents run? There are a few modes. You can run an agent locally against your own working directory, or run it in Cursor’s cloud on a dedicated VM. Cursor also supports self-hosted cloud agents, where the VM lives in your own infrastructure but still plugs into the same agent experience. That matters for teams that need internal network access or can’t let code and artifacts leave their environment. (cursor.com) ### What can the cloud version do? The cloud setup is the part that makes this feel operational, not just experimental. Each agent gets its own isolated VM, a cloned repo, and a configured dev environment. The run can keep going if your laptop sleeps or your network drops. You can reconnect later, inspect progress in Cursor’s Agents Window or web app, and let the agent push a branch or open a pull request when it finishes. (cursor.com) ### What tooling comes with it? This is where Cursor is trying to preserve the “full agent” feel outside the editor. The SDK works with skills, hooks, and subagents. Skills let agents pick up repo-specific capabilities from `.cursor/skills/`. Hooks let teams observe or control the agent loop with `.cursor/hooks.json`. Subagents let a main agent delegate named subtasks to specialized helpers with their own prompts and models. (cursor.com) ### How does the API look under the hood? Cursor also reworked its Cloud Agents API around durable agents and per-prompt runs. That sounds wonky, but the practical effect is useful: follow-ups, status checks, cancellation, and streaming are tied to a run, not bolted on as an afterthought. It also added SSE streaming, reconnect support through `Last-Event-ID`, lifecycle controls like archive and delete, and more standardized error shapes. (cursor.com) ### How is it priced? Not as a separate enterprise mystery box. Cursor says SDK runs use the same standard token-based consumption pricing, request pools, and privacy-mode rules as runs from the IDE and Cloud Agents. Spend shows up in the usage dashboard under the SDK tag. That makes the product feel less like a demo API and more like a metered extension of Cursor’s existing platform. (cursor.com)pers? Before this, Cursor was mainly a place where a human sat with an agent. Now it’s also a place where software can call an agent directly. That shift matters. Once an agent can be triggered from CI, a script, or a backend service — and can safely run in its own environment with billing, streaming, and PR creation built in — it stops being just an assistant. It starts looking like programmable labor for software teams. (cursor.com) ### Bottom line? Cursor didn’t invent coding agents this week. What it did do is package its own agent stack into something developers can embed, automate, and pay for like infrastructure. That’s the interesting part — not the TypeScript wrapper, but the idea that the editor agent is escaping the editor. (cursor.com)