GitHub Copilot computer use for QA is now in public preview in Copilot CLI and the GitHub Copilot app on Windows and macOS. Announced by GitHub on October 1, 2026, the capability lets Copilot interact with desktop apps through accessible content and visual context: it can click, type, scroll, drag, press keys, and move through workflows that may not expose an API, CLI, or MCP integration.
That makes the release relevant to QA engineers who still need to validate legacy clients, packaged desktop software, administrative consoles, or other GUI-only flows. It is a public preview, however, so it should be evaluated as a supervised assistant—not treated as a replacement for a deterministic regression suite.
What GitHub announced
Computer use is disabled by default. In Copilot CLI, enable it with /computer on; use /computer show to inspect status and /computer off to turn it back off. In the Copilot app, enable it under Settings > Computer Use, or use the same command. Copilot asks for approval before controlling an app, and organization-managed settings can disable the feature.
GitHub says the feature is intended for tasks where a visual interface is necessary. Its documentation specifically advises using an API, MCP server, terminal command, filesystem tool, or dedicated browser tool when one can complete the job directly, because those options generally produce more structured and predictable results.
Why this matters for QA engineers
Computer-use agents add a practical exploration layer around GUI-only systems. A tester can ask Copilot to navigate a known scenario, capture the outcome, and help identify where the UI diverges from an expected workflow. That can shorten exploratory testing and assist with reproducing a reported issue across several desktop applications.
- Legacy coverage: explore flows in software that has no stable automation interface.
- Reproduction support: walk through a clearly scoped bug report while a tester watches the tool activity.
- Cross-app checks: validate handoffs such as a desktop export opening in a spreadsheet or a document editor.
- Evidence review: compare the final visible state with an acceptance checklist before filing or closing a defect.
A safe preview test plan
Start with a disposable test account and a non-production environment. Choose a small, reversible workflow: create a draft record, edit a non-sensitive field, verify a confirmation message, then remove the test data. Give the agent the app name, expected outcome, and explicit boundaries.
Open the staging desktop client and sign in with the QA account.
Create a draft order for the test customer.
Do not submit, send, delete, or change any settings.
Stop after the confirmation banner is visible and report the steps taken.Record whether approval was requested, whether the selected controls were correct, the final UI state, elapsed time, and any retry or unexpected action. Repeat the same case with changed window size, a delayed response, an ambiguous label, and an error dialog. Those variations matter because GitHub notes that application versions, operating systems, window state, timing, and dynamic controls can cause an agent to select the wrong target, repeat an action, or fail to continue.
Controls to keep in place
- Keep human approval enabled for apps that can modify data or affect other people.
- Avoid Always allow for sensitive, financial, production, or enterprise-administration tools.
- Verify what is visible on screen before sharing it as agent context; windows can expose personal or confidential information.
- Use an API or conventional test automation where it is available and reliable.
- Review the end state before accepting an agent-reported pass or failure.
Copilot computer use is promising for supervised GUI exploration, but the important QA question is not whether the agent can click through a happy path. It is whether its permissions, targeting behavior, recovery, and evidence are dependable enough for the specific risk level of the workflow. Pilot it with narrow scenarios, reproducible test data, and a human in the loop.