GitHub opened public preview access to Computer Use on October 1, 2026, a feature that lets GitHub Copilot operate desktop applications directly: clicking, typing, navigating menus, without going through an API or a command line. The announcement targets a blind spot that automation has never quite reached: professional software with no API, no CLI, and no MCP server.
Built for apps that have no API
Computer Use ships in the GitHub Copilot CLI and in the Copilot desktop app, on macOS and Windows. In practice, the agent reads an app’s accessible content and visual context, then acts: clicking controls, entering and editing text, pressing keys, scrolling, dragging, and navigating across applications to carry out an entire workflow.
GitHub points to concrete use cases: filling out an expense report in a closed accounting tool, summarizing notifications scattered across several apps, updating a presentation, or moving data between tools that were never built to talk to each other. It extends the same logic already at work in editorial agents running in production, this time applied to graphical interfaces that expose nothing to machines.
What Copilot can do, and what stays guarded
The feature does not activate silently. Copilot asks for approval before taking control of an app, and users can review or reset the list of already-approved applications. On macOS, setup walks users through the Accessibility and Screen Recording system permissions required for GUI control. Organizations can disable the feature through managed settings, making it a global switch rather than a per-device choice.
| Automation approach | App-side requirement | Main limitation |
|---|---|---|
| API or webhook | Documented, stable endpoint | Missing from most closed line-of-business software |
| MCP server | Integration built for the specific tool | Upfront implementation work, not always justified |
| CLI script | Exposed command-line interface | Absent from strictly graphical software |
| Computer Use | None: reads the screen and accessible controls | Depends on the target interface staying visually stable |
With Computer Use, GitHub Copilot no longer just reads code: it clicks, types and navigates any desktop application, whether it exposes an API or not.
Turning it on
In the CLI, /computer on enables the mode, /computer show displays its status, and /computer off turns it off. In the desktop app, the same switch lives under Settings, then Computer Use, or directly via /computer on in a session.
# In the GitHub Copilot CLI
/computer on # enable desktop app control
/computer show # show status and approved applications
/computer off # disable the feature
The feature is still in public preview: GitHub recommends limiting it to non-critical applications until a reliability track record builds up, and reviewing every action taken on sensitive data.
An alternative to MCP, not a replacement
Computer Use does not compete with integrations built around the MCP protocol: it covers the ground they do not reach. An internal tool with no programmable integration at all remains, for now, out of reach for classic agents driven through a CLI or an API. GitHub is betting that visual automation, long confined to fragile UI-testing scripts, becomes reliable enough for everyday professional use once handed to a model that reads screen context instead of fixed coordinates.
The takeaway
Computer Use opens up ground previously reserved for specialized RPA tools: operating API-less software by reading the screen and interacting directly with its controls. The feature is in public preview, gated behind explicit approval, and can be disabled at the organization level, which makes it a cautious addition rather than a wholesale shift of automation toward agents.
On paper, Computer Use solves a real problem: most of the line-of-business tools teams use daily were never built to talk to a machine. But driving a graphical interface by reading the screen is still a bet on reliability, not a certainty: one layout change, one unexpected confirmation dialog, and the automation breaks silently. Before wiring Computer Use into anything touching invoicing or customer data, it’s worth running it for several weeks on low-stakes tasks first, and keeping an eye on every session while GitHub itself still calls it a preview. — Simon Janvier
Further reading: the official announcement on the GitHub changelog.
