Type /computer-use and your agent sees the screen, reads each app's accessibility tree and drives real apps on Windows and macOS with your mouse and keyboard. Esc always stops it.
/computer-use let Gemini 4 Hydrogen paint this in Microsoft Paint. One prompt in Antigravity. The agent drove Paint through a stroke plan scripted from the photo, shown 200× faster. The bar and glow are drawn in, since they're hidden from screen recordings by default.
Hover over the painting to see the strokes. Every one was a real drag of the mouse in Paint.


On a phone, tap an image to open it full size.

One skill and a small local service. It uses the model you already chose in Antigravity: no API key, no second LLM, no MCP server.
Reads the window's accessibility tree, the names, roles and values of its controls, and takes a screenshot only when a task needs eyes.
Clicks, types, drags and fills in dialogs with the real mouse and keyboard, on the elements it just observed.
Checks a fresh view after each step. An input whose outcome is uncertain is never sent twice.
Click, select, toggle and set values on buttons, fields and lists by name.
Click, double-click, right-click, scroll and draw strokes through hundreds of points on a canvas.
Unicode text, shortcuts, and paste that puts your clipboard back afterwards.
Launch, find and focus windows, including each app's own file dialogs.
plan sends many steps at once, so a 30-turn task drops to about 9.
Signed releases, automatic updates you can switch to manual, and rollback.
A small bar always says what the agent is doing right now, and it changes the moment the state does.
Soft light around the window the agent is working in, so you always know where it is. It's left out of the agent's own screenshots.

Nothing runs without your OK. Terminals and code editors are click-only, so it can't type commands into them.
Esc or Ctrl Alt F12 stops it at once and lets go of every key and button. Sign-ins and admin prompts are always left to you.
Not a sandbox and not a browser extension: the agent works in the same desktop apps you do.
The portrait above: exact colours set through Paint's own colour editor, then painted as long strokes with the real mouse.
Open the app, pick a photo in its file dialog and run a 4× upscale. Batching steps cut the turns the model needs.
Windows 10 and 11, x64. One-line install in PowerShell; uses the .NET Framework already on Windows.
macOS 14 or later on Apple silicon or Intel. You grant Accessibility and Screen Recording once.
No. It uses the model you already chose in Antigravity. There's no second LLM, no API key and no MCP server to set up.
It is built and tested for Google Antigravity. The service is driven by a plain command-line tool, so other agents that can run shell commands can call it, but only Antigravity is supported today.
It only acts in apps you approve, terminals and code editors are click-only, and Esc stops it at once and releases every key and button. Your own mouse and keyboard never pause it, so stopping is always one key away.
Most of the time in a long task is the model thinking between steps. plan removes much of that by sending many steps in one command.
Yes, MIT licensed. It is an independent project, not affiliated with or endorsed by Google.
Windows, in PowerShell
irm https://raw.githubusercontent.com/shayann07/antigravity-computer-use/main/install.ps1 | iexmacOS (Apple silicon or Intel)
Download the macos-arm64 or macos-x64 archive from the latest release, then run install.command and Setup.Mac.Permissions.command. Mac guide
Then try
/computer-use
Open Paint, draw a red circle in the middle, save it to my Desktop as circle.png.