Cursor releases TypeScript SDK for coding agents with cloud sandboxes and token-based pricing

Cursor released its TypeScript SDK for coding agents on April 29, 2026, bringing agents out of the editor and into code and automation. The same agents that work in Cursor's desktop app, CLI, and web version can now be spun up with just a few lines of TypeScript — locally, in the cloud, or on your own workers.

One SDK Everywhere

The new `@cursor/sdk` package gives developers programmatic access to the same runtime, harness, and models that power Cursor's agents. Essentially, the company is turning its editor's AI stack into an infrastructure layer: you can create an agent from a Node.js app, script, or pipeline and then interact with it as a service. Cursor emphasizes that the SDK is already available to all users in public beta, and getting started amounts to a standard `npm install` and calling `Agent.create()`.

The bet isn't just on local execution. Cloud sessions for the SDK run on the same platform as Cursor Cloud Agents: each agent gets its own virtual machine with hard sandboxing, a cloned repository, and a ready-made development environment. Such an agent can continue a task even if your laptop goes to sleep or the connection drops. Developers can stream the agent's progress, reconnect to the session later, and after completion receive a ready branch, pull request, demo, or screenshots.

Tools Inside the Agent

The release's strength lies in the fact that Cursor isn't offering a "lightweight" API version but is handing over nearly its entire agent circuit. The SDK inherits codebase indexing, semantic search and fast grep, can connect MCP servers via stdio or HTTP, automatically picks up skills from the repository, and supports hooks for controlling the agent's work cycle. Subagents are highlighted separately: the main agent can delegate subtasks to specialized assistants with their own prompts and models. Plus, developers can route tasks to any model that Cursor supports, including its own Composer 2 coding model.

• Dedicated sandboxed VMs for cloud sessions
• Connecting MCP via `.cursor/mcp.json` or inline config
• Hooks via `.cursor/hooks.json` for observing and extending the cycle
• Subagents with separate roles, prompts, and models
• Auto-creating branches and PRs after task completion

There's a separate emphasis on usage economics. In the announcement, Cursor writes that the SDK is priced according to the standard token-based consumption model. On the pricing page, the company clarifies that the plan already includes a certain volume of model usage, and beyond that, on-demand billing applies based on actual consumption. For teams, this is an important signal: the SDK isn't being sold as a separate enterprise product with a new license, but as a way to move your current agent infrastructure into CI/CD, internal services, and user applications without rebuilding the entire system from scratch.

Scenarios for Teams

Cursor presents the release not as a toy for demos but as a tool for production scenarios. In the company's examples, programmatic agents are launched directly from CI/CD to summarize changes, hunt for root causes of CI failures, and update pull requests with ready-made fixes. Another class of use cases involves internal platforms where non-engineering teams can ask questions about product data without writing code. An even more ambitious scenario is embedding the agent experience directly into client products, so the end user never leaves the main application interface.

"This is the path to your own agents in the same cloud runtime without managing VMs," as they described the release's idea at

Faire.

To lower the barrier to entry, Cursor has published a public cookbook repository with ready-made templates. It includes a minimal Quickstart for a local agent, a prototyping web app for running cloud agents, a kanban board where you can hand a card to an agent and get back a PR, and a lightweight CLI for starting agents from the terminal. This set shows that the company is aiming not just at the editor but at the layer of tools around development — from interfaces for managers to automatic work queues.

What This Means

Cursor is taking another step from an AI editor to a platform for autonomous development. If coding agents used to live mostly inside the IDE, they can now be embedded into company processes as a regular software component. For the market, this reinforces the trend of "agents as infrastructure": winners will be those players who offer not just a smart model but also a safe runtime, control, integrations, and clear economics for running in production.