Cursor Computer Use: what it is and how to use it
2026-10-26 · 4 min read
Cursor Computer Use is a new tool that lets Cursor's coding agent click and type in desktop apps on your own computer, not just in a browser. To use it, update Cursor, then let the agent pick it when a task needs it, or type @ in your prompt and choose Computer Use.
The posts that announced it
On Friday, @cursor_ai posted a short video with the line "Cursor can now control your computer." The demo used the iOS Simulator: the agent builds a feature, opens the simulator, and taps through the app to check its own work. By Monday morning the post had more than 6,000 likes and over 400 quote posts.
A follow-up in the same thread added the details. Computer Use lets agents "click and type across any app on your computer, not just the browser." Cursor says the agent uses it on its own when a task calls for it, and you can force it by @-mentioning Computer Use. The only setup instruction was to update Cursor from its download page.
@n2parko, whose bio lists product at SpaceX AI, quoted the announcement with "local computer use has arrived." Local is the part that's new. Cursor has belonged to SpaceX since August (the deal that led OpenAI to pull its models from Cursor), and this release puts the agent on your actual machine.
What changed from before
Cursor agents could already use a computer, as long as it was in the cloud. A February Cursor blog post described cloud agents that each get an isolated virtual machine, test their changes there, and send back videos and screenshots. Cursor's docs later added computer use for self-hosted workers on Mac and Linux.
| Where it runs | What the agent controls | How you start it |
|---|---|---|
| Cursor desktop app (new) | Apps on your own computer | Update Cursor, then @-mention Computer Use or let the agent pick it |
| Cloud agents | Its own isolated VM | Send a task from cursor.com/agents |
| Self-hosted worker (Mac, Linux) | The machine running the worker | agent worker --computer-use start |
According to AlphaSignal's coverage, Cursor announced no pricing change with the release. Several write-ups also pointed out that release notes and setup docs for the local desktop version weren't out when it shipped. So some of what follows comes from Cursor's self-hosted docs and early third-party walkthroughs.
How to try it
In the desktop app, update Cursor and open an Agent chat. Type @ and pick Computer Use from the menu, or describe a task that obviously needs a desktop app. On a Mac you'll need to grant two permissions in System Settings, under Privacy & Security: Accessibility, for the mouse and keyboard, and Screen Recording, for screenshots. Cursor's self-hosted docs say both go to a helper app called Cursor Computer Use, not to Cursor or Terminal, and early walkthroughs report the same helper on the desktop app.
Start with a small task whose result you can check. Some prompts along the lines of Cursor's demo:
- "Build the signup screen, run it in the iOS Simulator, tap through the flow and screenshot each step."
- "Open the app, try to submit the contact form with an empty email field, and tell me what happens."
- "Launch the desktop build and confirm the settings window opens without errors."
On a self-hosted Mac or Linux machine, Cursor's computer use docs say to run agent update and then start the worker with agent worker --computer-use start. Running agent worker debug confirms the helper is installed but doesn't check permissions, so Cursor suggests a real test: ask the agent to open a browser, visit a page and take a screenshot. On Linux you have to install packages such as xdotool and ffmpeg yourself. Enterprise admins can switch the feature off under MCP Configuration in team settings. It's on by default.
My read
Testing is where this pays off first. An agent that can open the app it just built and poke at it will catch bugs that unit tests miss, and mobile and desktop teams have been doing that part by hand.
The catch is trust. Whatever is on your screen during a task goes to the model, and an agent that can click can also send, delete or submit. Anthropic's report on Claude agents taking unintended actions shows what that looks like when agents touch real systems. I'd close sensitive apps, keep tasks narrow, and consider a separate macOS user for agent sessions.
For a small business, the takeaway is that agents now work in desktop apps too, and that's where a lot of office work happens. I'd hand them work you'll review afterward and keep a person on anything that can't be undone. If you want help sorting out which of your processes an agent should run, New Face Design's free process audit is a reasonable place to start.