GitHub has added computer use to Copilot, letting the AI assistant operate desktop applications by reading what is on screen, clicking, typing and scrolling, the company said Thursday in its changelog. The feature is in public preview in the GitHub Copilot CLI and in the GitHub Copilot app for macOS and Windows.

What Happened

With computer use switched on, Copilot can read the accessible content of an application window and act on it the way a person would: clicking buttons and other controls, entering text, pressing keys, scrolling, dragging and stepping through multi-screen workflows. GitHub pitched the capability at software that coding agents have so far been unable to reach, such as older line-of-business programs, tools that only offer a graphical interface and systems with no API or command-line access.

The feature is off by default. In the Copilot CLI, users turn it on with the /computer on command, check its state with /computer show and disable it with /computer off. In the Copilot app it sits under Settings, in a Computer Use section, and the same slash command also works there.

RelatedClaude Code Desktop Preview Flaw Could Have Exposed Developers’ Cloud Keys, Researchers Say

GitHub gave examples of the kinds of jobs it has in mind, including summarizing notifications in a browser, editing a presentation and moving information from one desktop application to another. The company said results are best when users describe the outcome they want, name the applications involved and spell out any constraints.

Why It Matters

Until now Copilot's agent features have mostly worked through files, terminals and GitHub's own services. Letting it act on any window on a developer's machine extends its reach into the parts of enterprise workflows that never got a programmatic interface, which can include internal admin consoles, desktop testing tools and vendor software.

It also widens what an agent can do if it misreads an instruction or is steered by malicious content on screen. GitHub's controls reflect that. Copilot must ask for permission before it interacts with a given application, and users can later review the list of apps they have authorized for automatic access and reset it. On macOS, setup walks users through granting the operating system's Accessibility and Screen Recording permissions, which any tool that reads and controls other apps requires.

For companies, the most important control may be the administrative one: GitHub said organization-managed settings can disable computer use. That gives security teams a way to keep the feature off fleet-wide while they assess it.

Context

GitHub is following a path other AI developers have already taken. Anthropic introduced a computer-use capability for Claude in October 2024, and OpenAI launched its Operator browser agent in January 2025. Those tools established screen-reading and simulated input as a fallback for systems that lack APIs, along with the accompanying concern that an agent with mouse and keyboard control can be manipulated by what it sees.

GitHub's version is notable for where it lives. Placing computer use inside the Copilot CLI and desktop app puts it in front of developers who already use Copilot for coding tasks, rather than in a separate product.

What to Watch

GitHub did not say when computer use will become generally available, whether it will be limited to particular Copilot plans or how it will be priced. The changelog also did not list known limitations beyond the admin controls. How enterprises configure the organization-level setting during the preview, and whether GitHub adds finer-grained policies such as allow lists of approved applications, will indicate how widely the feature is used in managed environments.