Design & engineering / 2026
macOS Computer Use
Giving agents a precise way to inspect and operate Mac interfaces.
A Model Context Protocol server with a native Swift core. It connects agents to macOS screenshots, Vision OCR, and Accessibility controls, with explicit targeting and permission boundaries.
A window is a real boundary.
A supplied window ID is authoritative. Invalid targets and app/window mismatches produce errors, rather than sending an action to an unrelated window. Semantic actions can work in the background without taking over the foreground app.
Use the capability the task needs.
Reading Accessibility controls and capturing the screen require different permissions. The server exposes a preflight so an agent can see what is available. Global input and arbitrary AppleScript are disabled in the default configuration.
Make actions observable.
Actions report how they were delivered and whether the result was verified. A separate cursor overlay makes the agent’s target visible without moving the person’s own pointer.
Swift / Node.js / MCP / Vision / Accessibility
Pixelator
Design tools