GitHub Copilot Computer Use Tutorial: Automate Desktop Apps the Terminal Could Not Reach

Introduction

On October 1, 2026, GitHub moved computer use into public preview in the Copilot CLI and the GitHub Copilot app on macOS and Windows. With it, Copilot can read an application’s accessible content and visual context, then click controls, enter and edit text, press keys, scroll, drag, and navigate workflows across applications on your behalf. It targets the automation blind spot every team knows: legacy and GUI-only software with no API, no CLI, and no MCP integration. This tutorial walks through enabling computer use, granting the required macOS system permissions, configuring approval controls, running a first desktop task, and resetting access when needed. It is written for developers and operators who already use Copilot and want to extend it to the software their terminal could never reach.

Quick answer

Enable computer use with /computer on in the Copilot CLI or Settings to Computer Use in the Copilot app, grant macOS Accessibility and Screen Recording permissions, approve each app at the prompt, then describe the outcome and app to automate.

Evidence

Computer use is landing on a platform with enormous scale. GitHub Copilot reached 50 million users in July 2026, Microsoft CEO Satya Nadella said on the company’s fourth-quarter FY26 earnings call, more than tripling from over 15 million in April 2025, according to Second Talent’s 2026 GitHub Copilot statistics roundup sourced from Microsoft earnings calls. Paid adoption is growing just as fast: Microsoft reported 4.7 million paid GitHub Copilot subscribers as of January 2026, up 75 percent year over year on its FY26 Q2 earnings call, reported by Panto AI’s statistics roundup. Enterprise penetration explains the urgency of controls like approvals and admin restrictions: more than 90 percent of the Fortune 100 use GitHub Copilot, per Microsoft’s Q4 FY25 earnings update reported by ITBrief. As of October 2026.

GitHub’s October 1, 2026 changelog states the control model in plain terms: “Copilot asks for approval before controlling an app.” That approval-first design, plus organization-managed settings that can disable the feature, is what makes a public preview of desktop control governable enough to try. Full setup and behavior details are in GitHub’s official computer use documentation, and the launch scope is pinned down in GitHub’s October 1 changelog entry, which covers the CLI, the app, and the supported platforms.

Prerequisites

• A GitHub account with a Copilot plan. GitHub documents that the Copilot CLI is available with all Copilot plans. If you receive Copilot through an organization, the Copilot CLI policy must be enabled in the organization’s settings.
• A macOS or Windows machine. Computer use is documented for local sessions on macOS and Windows only; Linux sessions are not covered by the current setup flow.
• The Copilot CLI installed, or the GitHub Copilot app installed for the graphical setup path.
• On macOS, administrator rights to grant the Accessibility and Screen Recording system permissions.
• For enterprise users: confirm your organization has not disabled computer use through managed settings before you invest time in a rollout plan.

Step-by-step guide

Every command and UI label below comes from GitHub’s current documentation for computer use, verified against the live docs page before writing. Computer use needs a macOS or Windows machine with the Copilot app or CLI, so treat this as the documented flow rather than a lab-verified run.

Step 1: Install or update the Copilot CLI

GitHub’s getting-started docs list three install routes. Pick the one for your platform; the npm route needs Node.js 22 or later:

# Cross-platform (npm, requires Node.js 22+)
npm install -g @github/copilot

# macOS (Homebrew)
brew install copilot-cli

# Windows (WinGet)
winget install GitHub.Copilot

Confirm the install with:

copilot --version

If you use the GitHub Copilot app instead of the CLI, skip to Step 3; the app path and the CLI path both toggle the same feature.

Step 2: Enable computer use in the CLI

Start an interactive session with copilot, then enable the feature with the slash command:

/computer on

Check its current state at any time:

/computer show

/computer show reports whether computer use is on, along with the current plugin and MCP status. To disable it later, run /computer off. These three commands are documented in GitHub’s launch changelog and are the fastest way to toggle the feature without opening settings.

Step 3: Enable computer use in the Copilot app

In the GitHub Copilot app, follow the settings path exactly as documented:

1. Open the settings for the GitHub Copilot app.
2. In the sidebar, select Computer Use.
3. Under Enable Computer Use, turn on Enable Computer Use.

You can also type /computer on inside an app session; it flips the same switch.

Step 4: Grant the macOS system permissions

On macOS, the OS gates what an agent can do to other applications, so GitHub’s docs add a Prerequisites step inside the Computer Use settings page:

1. Under Prerequisites, grant the Accessibility permission. This allows computer use to interact with application controls such as buttons, menus, and text fields.
2. Grant the Screen Recording permission. This allows computer use to inspect application windows when it needs visual context.
3. If a permission change is not detected, click Check again. This is GitHub’s documented recovery step for the permissions UI lagging behind System Settings.

Both permissions must show as granted before Copilot can interact with desktop applications.

Step 5: Configure approval controls

Computer use follows the app’s Tool Permissions setting, which you can review by opening the app settings and selecting Sessions. When approval is required, GitHub’s docs describe the exact prompt behavior: check that the requested application and the requested action match your prompt, then choose Allow to grant access for the current computer-use session, Always allow to save the approval for future sessions, or decline or cancel the request.

Treat Always allow as a real security decision. GitHub documents that an Always allow granted in the Copilot app or the Copilot CLI is saved and applies to both surfaces on the same computer, so one careless click extends trust to every future session on that machine. Start with per-session approvals and only promote a specific app to Always allow once you have watched it behave.

Step 6: Run your first desktop task

Computer use works best when you describe the outcome you want, the applications involved, and any important constraints. GitHub’s docs give a canonical prompt template for a safe first run:

Open APP_NAME and summarize the status information shown in the
main window. Do not change any values or submit any forms.

Replace APP_NAME with an application installed on your computer. Notice the two safety properties of this template: it names one app, and it forbids changes and form submissions. GitHub’s own launch demo follows the same pattern with a real task, an expense-report workflow in Safari, where the agent reads screens and clicks through while the user watches the approval prompts.

For your first week, keep to read-only tasks: summarize what is on screen, find a record in a legacy client, compare values across two windows. Graduate to tasks that change state only after you have seen the approval prompt fire correctly and stopped an operation with Esc at least once.

Step 7: Review, reset, or disable access

Auditing is part of the feature, not an afterthought:

1. Open the settings for the GitHub Copilot app.
2. In the sidebar, select Computer Use.
3. Under Always allowed apps, click the delete icon next to the application you want to remove.

GitHub documents two important consequences here. Removing an application deletes its saved approval for future sessions in both the app and the CLI, but it does not revoke access already granted in the running session. To stop an active operation immediately, click Stop or press Esc; end the session to revoke application access granted to that session.

To turn the feature off entirely, either run /computer off in a session or open the app settings, select Computer Use, and turn off Enable Computer Use.

One more boundary to keep in mind: computer use and local sandboxing are separate switches, and both ship turned off. Enabling computer use gives the agent reach into your desktop applications; it does not fence what the agent’s shell commands can read or write. Keep the CLI’s own permission model tight alongside it.

Troubleshooting

Problem 1: The Computer Use settings page is not available.

This is the first documented failure mode, and it is usually not a bug. GitHub’s docs say to confirm that the feature is available for your account and operating system. Check that you are on macOS or Windows with a current Copilot app or CLI, and that your organization has not disabled computer use through managed settings. If any of those conditions fails, the page will not appear.

Problem 2: On macOS, a permission change is not detected.

System Settings and the Copilot app can disagree for a few seconds after you flip a toggle. GitHub’s documented fix is to click Check again under Prerequisites. If the permission still does not register, close and reopen the Copilot app settings and re-check; the Accessibility or Screen Recording entry should eventually show the new state.

Problem 3: Copilot cannot interact with applications at all on macOS.

Verify both prerequisites independently. GitHub documents that computer use needs Accessibility and Screen Recording to show Granted under Prerequisites. Accessibility alone lets it see controls but the Screen Recording permission is what allows inspecting application windows for visual context; with only one granted, interaction fails in exactly the confusing way you would expect.

Problem 4: Copilot asks for approval on every single action.

That is the Tool Permissions setting doing its job, not a defect. Open the app settings, select Sessions, and review the setting that controls whether Copilot asks for approval before controlling an application. You can either live with per-action approval for sensitive apps or promote a trusted app to Always allow, accepting that the saved approval applies to both the app and the CLI on that computer.

Problem 5: You granted Always allow to the wrong app.

Remove it under Always allowed apps with the delete icon, which clears saved approvals for future sessions in both surfaces. Remember the documented caveat: this does not revoke the running session. Click Stop or press Esc to halt the current operation and end the session to revoke the access it was already granted.

Problem 6: /computer on appears to do nothing.

First run /computer show to confirm the state. If the feature will not enable, the likely cause is organizational: organization-managed settings can disable computer use, and the Copilot CLI policy must be enabled in the organization’s settings for org-provided Copilot. Individual users should also verify they are on a supported OS, since the feature is documented for local sessions on macOS and Windows only.

FAQ

Is computer use production-ready?

No. GitHub ships it as a public preview and marks the docs page as subject to change. It is the right stage for controlled experiments on non-sensitive workflows, with per-session approvals, not for unattended automation of anything that moves money or data.

When should I reach for computer use instead of an MCP server?

When there is no better interface. Our MCP server tutorial shows how to give agents a structured, auditable tool for apps that support it, and APIs and CLIs are always preferable to screen driving. Computer use exists for the long tail of legacy and GUI-only software where none of those exist. For the broader question of which automation platform to standardize on, see our n8n vs Make comparison for AI agents.

How does this compare to agent orchestration approaches like Claude’s dynamic workflows?

They solve different problems at different layers. Claude’s dynamic workflows orchestrate up to a thousand agents on a managed platform, while computer use gives one Copilot session hands on one desktop. You can think of computer use as a new tool in the agent’s belt and orchestration as the system that decides when to use it.

Can computer use run while I am away from the computer?

Nothing in GitHub’s documentation supports unattended operation. The approval model assumes you are present to check that the requested application and action match your prompt, and Stop / Esc assume you are watching. Plan for supervised use.

Does enabling computer use in the CLI also enable it in the app?

The setting is shared, and so are saved approvals: an Always allow granted in either surface applies to both on the same computer. Review the Always allowed apps list in the app’s Computer Use settings to see the full picture, and see our Claude Code subagent effort levels guide for related thinking about scoping what agents are allowed to do.

Conclusion

GitHub Copilot computer use extends the agent from the terminal to the desktop, opening up the GUI-only software that APIs, CLIs, and MCP servers cannot reach. The public preview pairs that reach with an approval-first control model: explicit enablement, per-app approval prompts, a reviewable Always allowed apps list, and organization-level disablement. Enable it with /computer on, grant the macOS permissions, start with read-only tasks, and audit your approvals regularly. As a next step, our Claude dynamic workflows tutorial shows how agent orchestration scales the other direction, from one desktop to fleets of agents.

Frequently asked questions