Cursor SDK adds custom tools and auto-review

Cursor's June 4 SDK release is a real workflow upgrade for teams building agents into scripts, CI, and internal tooling. Instead of forcing developers to stand up separate MCP servers or accept all local tool calls blindly, Cursor now gives SDK users more direct control over tools, persistence, and approvals. The release matters because it pushes the SDK from experimentation toward production readiness. Custom tools, nested subagents, and safer auto-review make it easier to build longer-running agent systems without giving up observability or control.

Key Features or Updates

The update adds `local.customTools`, letting developers expose their own functions directly to agents through Cursor's built-in MCP path. It also adds `local.autoReview`, JSONL and custom local stores, nested subagents, and a batch of reliability fixes around streaming, checkpoints, and local `wait()` behavior.

Impact on Developers

This lowers the setup cost for custom agent workflows and makes headless runs less risky. Teams can now keep more of their agent logic inside application code, persist state in formats that fit their stack, and add review guardrails without building a parallel permissions layer.

How to use it

Upgrade the SDK with `npm install @cursor/sdk` or `pip install cursor-sdk`, then define custom functions through `local.customTools`. If you want approval logic in unattended runs, enable `local.autoReview` and tune `permissions.json` so the classifier can distinguish safe read-only calls from destructive operations.