GitHub Copilot computer use: what it does

GitHub Copilot is no longer limited to code and the terminal. On 1 October, GitHub launched a public preview of "computer use," which lets Copilot operate ordinary desktop apps the way a person would: reading what's on screen, clicking, typing, and scrolling.

Everything below is from GitHub's own announcement.

What's confirmed

What it's for

GitHub pitches it at software that has no other way in: older programs and graphical apps without an API, command-line interface, or MCP integration, the standard many AI tools now use to connect to apps. If the only way to use a program is to click through it, Copilot can now do the clicking.

GitHub's own advice is to describe the outcome you want, name the apps involved, and spell out any constraints, rather than giving step-by-step instructions.

The bigger picture

It lands two days after OpenAI announced an Agents API with computer use at DevDay. Agents that act on your behalf, inside real software, are where the big AI companies are putting their effort this autumn.

It's also the same capability behind this month's safety worries, from agents getting out of their tests to the FTC's new investigation. GitHub's approach, asking before it touches each app and letting organisations turn it off, is the sensible baseline. Whether it's enough is what a public preview is for.

Should you try it

If you're a developer who regularly clicks through the same tools, yes, carefully. Start with apps where a wrong click can't do much harm, and don't grant lasting access to anything that handles money, accounts, or production systems until you've seen how it behaves.