GitHub Copilot can now operate desktop apps on macOS and Windows

GitHub put computer use into public preview on October 1, 2026, for GitHub Copilot CLI and the GitHub Copilot app on macOS and Windows. With the feature on, Copilot can read what a desktop application shows, click its controls, type into it and move between apps to finish a task. GitHub announced the change in a short changelog entry, and its docs describe the permission model and the risks in more detail.

What computer use in GitHub Copilot adds

Computer use gives Copilot a way to act on a graphical interface. According to the GitHub changelog entry (opens in a new tab), the target is work in legacy and GUI-only software that has no API, command-line interface or MCP integration.

The concepts page for computer use (opens in a new tab) lists what the agent can do once the feature is enabled:

• read application content and visual context, through the operating system's accessibility tree or through screenshots when it needs visual context
• click controls
• enter and edit text
• press keys, scroll and drag items
• move through workflows that span several applications

Example tasks in the docs include summarizing information in a legacy desktop app, updating a presentation, entering data into GUI-only software and moving information between applications.

How to turn on computer use

Computer use is off by default. In Copilot CLI, the CLI guide for computer use (opens in a new tab) gives three commands for an interactive session:

/computer on
/computer show
/computer off

/computer on saves your preference and enables a bundled computer-use plugin. /computer show reports whether that plugin, its MCP server and the related skills are active. In the GitHub Copilot app, open Settings, select Computer Use and turn on Enable Computer Use, or type /computer on.

On macOS, Copilot walks you through granting two permissions: Accessibility, to interact with controls, and Screen Recording, to inspect windows when it needs visual context. The CLI guide says computer use runs in local sessions only.

GitHub's advice on prompts is to state the outcome you want, the apps involved and any constraints. The CLI guide's sample prompt asks Copilot to summarize an app's status window and explicitly forbids changing values or submitting forms.

Approvals and admin controls

Computer use follows the tool permission settings of the Copilot surface you are in. When Copilot asks to control an app, you can allow it for the current session, choose Always allow to save the approval, or decline. Saved approvals are stored locally and apply to both the CLI and the app on the same machine. The app shows the list of always-allowed applications, and removing one does not revoke access already granted in a running session.

Deny rules take precedence over automatic or saved approvals. To stop a run in the CLI, press Esc twice. In the app, click Stop or press Esc.

For organizations, enterprise administrators can switch computer use off through enterprise managed settings. A user's local setting cannot override that policy, and the CLI reports when the feature is blocked.

Who this helps

Computer use matters most for teams with internal tools or older desktop software that has no programmatic interface. The docs are direct about the trade-off. If an API, MCP server, terminal command, filesystem tool or browser tool can do the job, GitHub says that route gives more structured information and more predictable results. For a system you control, writing a small MCP server in Python is often the more reliable path than screen automation.

Limits, risks and open questions

GitHub's docs list several failure modes. Computer use can click the wrong control, type in the wrong place, or struggle with non-standard and dynamic controls. Changes in timing or window state can make it repeat an action or stop partway through.

The security warning is explicit. Ambiguous instructions or unexpected on-screen content can lead to unintended actions affecting your device, data or connected accounts, including personal, financial or enterprise systems. Windows may show sensitive information about other people, and anything visible becomes context for Copilot. GitHub recommends not choosing Always allow for apps that hold sensitive data or can take high-impact actions.

Several things are not stated in the changelog or the docs pages:

• which Copilot plans include computer use (the CLI itself is available on all Copilot plans)
• whether computer-use sessions cost more AI credits than other agent work
• which models drive the feature
• Linux support, since only macOS and Windows are named
• when or whether the feature leaves public preview, which GitHub says is subject to change