GitHub’s advice for its new Copilot feature is to try something else first
The New Stack Amanda Caswell
GitHub put Copilot in apps on Windows and macOS. But it says to use an API, terminal, or MCP tool first.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
GitHub has put computer use into public preview, giving Copilot CLI and the desktop app the ability to work inside apps on macOS and Windows. The agent can read what’s on screen and click, type, scroll, and drag. That even extends to older software with only a graphical interface, where there’s no API, no command-line tool, and no MCP integration to lean on.
The launch demo was an expense report in Safari, but GitHub is also pitching the feature for jobs like summarizing data in a legacy app, updating a presentation, entering information, and moving content between apps. Developers can trigger it from the terminal or through the Copilot app, which runs on Copilot CLI and arrived earlier this year as a rival to Claude Code and Codex.
GitHub is not pretending this should replace cleaner integrations. Its own guidance is to use direct tools first. If an API, MCP server, terminal command, filesystem tool, or dedicated browser tool can do the job, GitHub says those routes usually give more structured data and more predictable results than having Copilot poke around the desktop.
That caution shows up everywhere in the rollout. Enabling computer use in Copilot CLI turns on a bundled plugin with its own MCP server, and on macOS it also needs Accessibility permission plus Screen Recording permission when it needs visual context. Approvals can be saved, shared between the CLI and desktop app on the same machine, and overridden by deny rules or enterprise policy. GitHub also warns that timing changes, window changes, or messy interfaces can make the agent repeat actions, stall, hit the wrong control, or expose sensitive information visible on screen.
My take — AI-written commentary, not fact-checked reporting
This is the right amount of caution for a feature like this. Desktop agents are useful when you’re trapped in some ancient GUI that nobody wants to touch, but they’re also a lovely way to turn a small mistake into a very expensive mouse click. The industry keeps reaching for “agentic” everything; GitHub at least admits that boring tools still do the better job most of the time.
Read more about this at: The New Stack