Anthropic just shipped one of the most ambitious features in its consumer product: Claude can now take over your Mac and use it like a person would. Point, click, scroll, type, open apps, navigate browsers - all running on your actual desktop.
The feature is live today as a research preview in Claude Cowork (Anthropic's desktop agent product) and Claude Code, available to Pro ($20/month) and Max subscribers on macOS.
How It Actually Works
Claude doesn't just blindly grab your mouse. It follows a hierarchy: first, it tries purpose-built connectors to services like Slack or Google Calendar. When no connector exists for what you need, Claude falls back to direct screen control - literally watching your display via screenshots and manipulating your mouse and keyboard to get things done.
That means Claude can open a spreadsheet, fill in data, switch to your browser, look something up, paste results back - the kind of multi-app workflow that would normally take you twenty minutes of clicking around.
Before accessing any new application, Claude asks for explicit permission. You can stop it at any point. This isn't a "hand over the keys and hope for the best" situation.
Dispatch: Start Tasks From Your Phone
The more interesting piece might be Dispatch, a companion feature that lets you assign tasks to Claude from your phone. Your Mac needs to be awake and running the desktop app, but you don't need to be sitting in front of it. Tell Claude what to do from your couch, your commute, wherever - it handles the work on your computer.
This is the kind of setup that makes computer use actually practical rather than a party trick. The value isn't watching an AI click buttons on a screen you're already sitting at. It's having it do that work while you're doing something else.
Still Rough Around the Edges
Anthropic is being upfront that this is early. Computer use "won't always work perfectly," and complex tasks may need a retry or two. That tracks with what we've seen from every computer-use AI agent so far - the technology works impressively well on straightforward workflows but stumbles on anything with unexpected popups, CAPTCHAs, or ambiguous UI states.
Anthropic isn't alone in this race. Perplexity shipped its own computer control agent, and several startups are chasing the same idea. But Anthropic has a head start on the underlying model capability - Claude's computer use API has been available to developers since late 2024, and this consumer launch builds on over a year of iteration.
For now, macOS only. No word on Windows or Linux support for the consumer product, though developers can already use the computer use API on any platform through the Anthropic API.