GitHub announced computer use in public preview on October 1, 2026, bringing desktop-app control to GitHub Copilot CLI and the GitHub Copilot app on macOS and Windows. For builders working with legacy or GUI-only software, it adds a way to complete steps that lack an API or other direct integration.
According to GitHub’s capability documentation, Copilot can inspect application content through the operating system’s accessibility tree or screenshots, then click, type, press keys, scroll and drag. The useful decision is which steps need that access—and how long an application approval should last.
Use desktop control for the missing interface step
GitHub recommends a direct API, Model Context Protocol (MCP) server, terminal command, filesystem tool or dedicated browser tool when it can do the job: those routes typically offer more structured information and predictable results. GUI access fills a gap; it need not replace working integrations.
Our decision framework, based on the documented limitations:
Use a direct tool when it can complete the required step and expose a result you can check.
Consider an attended desktop trial when a necessary step is GUI-only, the visible information is appropriate for Copilot, and the change is reversible and independently checkable.
Keep the step manual when it involves sensitive accounts or consequential changes without adequate review or recovery. A short sequence of clicks is not necessarily a low-impact task.
For example, a team could prepare approved text with existing tools, use desktop control only to place it in a presentation editor, and inspect the saved slide before sharing. That is an illustrative design, not a tested Copilot workflow. The same interface-selection question appears in our Holo4 guide to routing computer-use work; this preview adds Copilot-specific permission decisions.
Check the account and permission mode before enabling
The feature is off by default. In an interactive CLI session, /computer show reports its status; /computer on enables it and saves the preference. Use /permissions show to inspect the active permission mode. These are documented CLI controls.
In the app, activation is under Settings → Computer Use → Enable Computer Use. Its Tool Permissions setting is under Sessions. On macOS, GitHub requires Accessibility permission for interaction and Screen Recording permission for visual inspection. The app setup guide describes both prerequisites.
Enterprise administrators can prevent enablement with features.computerUse set to false. A local preference cannot override that restriction. Setting it to true or leaving it unspecified permits local choice; it does not itself grant application access. See GitHub’s managed-settings reference.
Separate session access and saved approvals
Do not assume a fresh prompt precedes every click. GitHub says the active CLI permission mode or app Tool Permissions setting determines whether an application request prompts or is approved automatically. Explicit tool-deny rules take precedence. Its agent application card describes these permission boundaries.
When prompted, Allow grants access for the current computer-use session; Always allow saves an approval. Saved app approvals apply to both Copilot CLI and the Copilot app on the same computer. Switching clients does not reset that consent. The application-access guide distinguishes the following controls:
Control | Documented effect | Important boundary |
|---|---|---|
Allow | Grants application access for the current computer-use session. | Not an approval for just one click. |
Always allow | Saves an app approval for later sessions in CLI and app on that computer. | Later access can proceed without another app-level prompt. |
Remove an always-allowed app | Deletes its saved approval for future sessions in both clients. | Does not revoke access already granted in a running session. |
Stop / Esc | Interrupts the active operation: Esc twice in CLI; Stop or Esc in the app. | Interruption is not documented as undoing completed actions. |
End the session | Revokes the application access granted to that session. | Review persistent approvals separately. |
The CLI’s double-Esc shortcut is also specified in its usage guide. To withdraw access, stop the operation, end its session, then review saved approvals and the effective permission mode. Removing an approval does not promise a new prompt if that mode automatically approves requests. Use /computer off if computer use should be disabled, and /computer show to check the resulting status.
Review what the app reveals, not just what it changes
GitHub’s application card explicitly places desktop interaction outside the CLI’s default directory-scoped filesystem access. Starting in a small repository does not confine an approved desktop app to that repository. Review the signed-in account and available data as well as the intended edit.
GitHub warns that visible windows can contain sensitive information and advises against Always allow for sensitive or high-impact apps. A request to summarize a window may expose data even without changing it. Asking for no edits expresses intent; the reviewed computer-use docs do not establish a separate enforced read-only mode.
For an initial trial, our recommendation is session-limited access, authorized test data and no unrelated sensitive windows. These precautions are not a technical isolation guarantee. Define success before starting: the right destination, correct values, expected saved status and no unintended duplicate.
GitHub documents possible targeting mistakes and repeated actions when interfaces or timing change. If a save is ambiguous, inspect the destination before retrying. Check any external effects after cancellation; a stopped session is not proof that an earlier submission was reversed.
What remains unestablished
The setup documentation specifies local macOS and Windows sessions and tells users to check account availability. The reviewed pages do not establish a computer-use plan-entitlement matrix, minimum client builds or Linux, remote-session and cloud-agent support. Confirm availability in the intended account before planning a rollout.
GitHub’s general billing documentation bases AI-credit consumption on model and token usage. The reviewed sources do not establish a fixed price per desktop task or a preview-specific success rate. They therefore cannot support a cost-per-completed-workflow or productivity claim.
Reporting note: This AI-assisted guide is based on GitHub’s announcement and documentation checked on October 1, 2026. Recommendations are analysis of those sources. No Copilot desktop workflow was run or benchmarked.