Cursor SDK: Programmatic Agents via TypeScript

Cursor Opens Its Agent Runtime to Developers

Cursor launched a public beta TypeScript SDK on April 29, 2026, making the same runtime, harness, and models powering the Cursor desktop app, CLI, and web application directly accessible to developers through a single package: npm install @cursor/sdk. The release marks a significant shift in positioning — from an interactive IDE assistant to deployable agent infrastructure that teams can embed into their own systems.

What the SDK Enables

The core value proposition is straightforward: developers no longer need to build and maintain an entire agent stack from scratch. The Cursor SDK delivers out-of-the-box capabilities that would otherwise require substantial engineering investment.

Intelligent Context Management

Agents created with the SDK benefit from the same codebase indexing and semantic search capabilities that make the in-editor Cursor experience powerful. When tasked with making changes or summarizing a repository, agents automatically understand codebase structure without manual configuration.

MCP Server Integration

The SDK supports connecting agents to external tools and data sources via the Model Context Protocol (MCP), either through a .cursor/mcp.json configuration file or passed inline at call time. This lets agents tap into databases, APIs, and other services as part of their reasoning and execution flow.

Skills Auto-Loading

Any skills defined in a repository's .cursor/skills/ directory are automatically picked up by SDK-driven agents, enabling teams to codify repeatable workflows once and have agents execute them reliably across runs without re-configuration.

Hooks and Subagent Delegation

Developers can observe and extend agent behavior through hooks defined in .cursor/hooks.json. For more complex workflows, agents can delegate tasks to subagents with custom prompts and model assignments — enabling multi-agent orchestration patterns without custom orchestration infrastructure.

Deployment Flexibility

One of the SDK's most notable characteristics is the range of execution environments it supports.

Local execution is available for rapid iteration during development, keeping everything on the developer's own machine. Cloud execution on Cursor's managed infrastructure gives each agent its own dedicated VM with strong sandboxing, a clone of the target repository, and a fully configured development environment. Cloud agents support persistent sessions that survive network interruptions and can create pull requests, push branches, and attach screenshots or demos. Self-hosted workers let organizations run agent execution within their own network, satisfying strict data-residency or security requirements without giving up the SDK's higher-level capabilities.

Model Access and Pricing

The SDK supports all frontier models available in Cursor, with Composer 2 — Cursor's in-house coding model — recommended as the default. Cursor positions Composer 2 as delivering frontier-level coding performance at reduced cost compared to general-purpose alternatives. Billing follows standard token-based consumption pricing.

Getting Started and Sample Projects

The SDK is available at @cursor/sdk on npm. Cursor published a public cookbook repository on GitHub (accumulating 2,300+ stars at launch) with four starter projects: a minimal Node.js quickstart, a web-based prototyping tool, an agent-powered kanban board with automated PR workflows, and a lightweight CLI for spawning Cursor agents from the terminal. The Cursor SDK plugin in the Cursor marketplace can assist developers as they build.

Early Adopters

Teams at Rippling, Notion, Faire, and C3 AI were already using the SDK ahead of the public beta. George Jacob, Senior Engineering Manager at Faire, noted that Cursor's cloud infrastructure handles parallel agent execution without developers needing to manage VMs or memory limits directly.

Cloud Agents API Improvements

Alongside the SDK launch, Cursor updated the underlying Cloud Agents API with four structural improvements: a restructured API emphasizing durable agents with run-scoped controls, streaming support via SSE events with reconnection capabilities, explicit lifecycle management (archive, unarchive, delete), and standardized v1 API responses with structured error codes.