GitHub Copilot ships computer use public preview, can drive desktop apps and multi-app workflows

GitHub shipped a public preview of computer use on October 1 in the Copilot CLI and the GitHub Copilot desktop app for macOS and Windows. The feature lets Copilot read an app's accessible content and visual context, click controls, type or edit text, press keys, scroll, drag, and walk through multi-step workflows across apps; the launch demo shows Copilot completing an expense-report flow in Safari.

The preview targets legacy or GUI-only software that does not provide an API, command-line interface, or MCP integration. The feature is disabled by default; users turn it on with /computer on in the CLI, check status with /computer show, and turn it off with /computer off. The desktop toggle sits in Settings under Computer Use. On macOS, Copilot walks the user through the required Accessibility and Screen Recording permissions. Every action runs through an approval gate; users can allow an app for one session, mark it Always allow, or deny outright. Organization administrators can switch the feature off via managed-settings.json, and the org policy overrides local enablement.

GitHub's documentation warns that computer use can pick the wrong control or type into the wrong place on dynamic interfaces; ambiguous instructions plus unexpected on-screen content can trigger actions affecting the user's device, data, or connected accounts, and sensitive information visible on screen can end up in the model's context. To interrupt an in-progress action, press Esc Esc in the CLI or click Stop in the app. GitHub frames computer use as the fallback when an API, CLI, filesystem tool, or dedicated browser tool is unavailable; more direct integrations remain the recommended path because they give more predictable results. The feature is excluded from GitHub's default-enablement policy for Business and Enterprise plans, so it will not switch itself on for enterprise seats on the preview date. The launch accompanies dynamic workflows (multi-agent processes defined in code) released the same day. OpenAI and Anthropic have previously shipped similar computer-use capabilities in Codex, Claude Code, and Claude Cowork.