Cursor announced Thursday that its agents can now click and type in any app on your computer, not just a browser. Their example: an agent driving the iOS Simulator to test the feature it just built. Cursor says the agent will reach for Computer Use on its own when a task needs it, or you can @-mention it.
How it works on self-hosted machines
Cursor’s docs for self-hosted workers spell out the setup, and it’s opt-in all the way:
- Start the worker with
agent worker --computer-use start. It’s never enabled by the server. - macOS: the CLI installs a small helper app, Cursor Computer Use. You have to grant it Accessibility and Screen Recording yourself, and it needs a signed-in desktop session.
- Linux: it uses an X11 display. It can reuse one you run, or start its own TigerVNC/Xfce desktop on a headless box. You install xdotool, ffmpeg and the X11 bits yourself.
- Desktop sharing (
--share-desktop) is Linux-only and lets authorized viewers watch or take control. No inbound ports open. - Enterprise admins can turn Computer Use off in team settings.
agent update
agent worker --computer-use start
agent worker debugMy take
This is the missing piece for agents that build apps: they can finally check their own work in the real UI. But an agent that can click anything on your desktop can click anything on your desktop. I’d run this on a dedicated VM or a spare machine, not my daily driver, at least until I’ve watched it for a while. Same rule I use for every new bot capability: by hand first, then trust.
Have a blessed day! And Happy Hacking!
