How to use computer use
Let the agent work in the apps on your Windows PC, one app at a time and as far as you allow: View, Click or Full. How to turn it on, what you see while it works, what's new in 0.2.14, and which models do it best.
Computer use lets the agent work in the apps on your PC the way you would: it looks at a window, clicks, types and reads what changed. It's off in every chat until you turn it on, and each app needs your OK. It works on Windows for now.
Turn it on in a chat
- In the message box, press 🖥 Computer use and tick Let the agent use apps on this computer. It's for this chat only.
- Ask for something in an app: "Open Calculator and work out 17% of 2,340", or "Draw a house in Paint". The agent works in apps that are already open, or opens installed ones by their Start-menu name.
- The first time it wants an app, a card asks you. Once you allow it, the app shows in the 🖥 panel with how far the agent may go in it.
| Level | What the agent may do in that app |
|---|---|
| View | Look only: its windows, their text, a screenshot |
| Click | Look, and click buttons and points |
| Full | Look, click, type and send keys |
Change an app's level in the 🖥 panel at any time, or remove it. Apps stay open after the agent uses them, so they can be used again.
While it works
A glowing edge goes round the screen and a Stop button stays on top while the agent uses your desktop (or one of your Chrome tabs). Press Stop at any time to end it.
- System tools and admin shells are never used, whatever the level.
- What you named in your message is your go-ahead. Ask for Paint and it can use Paint; an app you didn't name asks first.
- Dialogs like "Replace it?" ask you first, so the agent can't overwrite a file on its own.
To keep an eye on it from QuorX, 👁 Follow the agent shows the page or terminal it's using. The new 🧭 button follows the running reply's steps instead, one chat at a time.
New in 0.2.14
- Freehand drawing: one held stroke through many points, so it can draw a curve in Paint rather than a row of dots.
- Smooth held mouse moves, for dragging.
- Up to 40 steps in one go, so a drawing doesn't pause after every move.
Which model to use
Computer use works best with a model that sees pictures, because the agent reads the screen from screenshots:
- a Claude Code model, through Apps on this PC;
- an OpenRouter model with image input.
It isn't available on Grok Build models, which read text only; the chat says so.