Claude Code can now control your Mac

Anthropic has introduced a new research preview feature that gives Claude Code direct control over a user’s computer on macOS. Inside the Claude desktop app, Claude can now take screenshots, move the mouse, use the keyboard, open apps, and complete tasks directly on the machine.

This pushes Claude beyond coding assistance and into desktop automation.

In Anthropic’s demo, Claude opens OBS and starts a recording by locating the correct button on screen, taking a screenshot of the interface, and clicking the right control. In another example, it opens Finder, finds a PDF in Downloads, and sends it through ClickUp by navigating the app step by step. The workflow is powered by repeated screenshots, on-screen interpretation, and UI actions.

The bigger unlock comes when this feature is paired with Dispatch, Anthropic’s remote tasking capability. That allows users to send Claude tasks from their phone while Claude executes them locally on their Mac. In practice, that means a user can be away from their desk and still have Claude find files, open apps, send messages, log information, or check workflows on their own machine.

For now, the feature is in research preview, available only on macOS, and limited to the Claude desktop app for Claude Code and Claude workspaces. It also requires system permissions such as accessibility access and screen recording, and Claude requests confirmation before sensitive actions like sending messages.

There are also clear limitations. Browser interaction is restricted for security reasons, which means Claude cannot yet fully click and type freely in Safari the same way it can in local desktop apps. For deeper browser automation, users still need separate tooling such as Chrome-based extensions or Playwright-style workflows.

Even with those constraints, this is a meaningful step forward. Many real-world workflows still depend on desktop software, internal tools, and legacy systems without clean APIs. By allowing Claude to operate those interfaces directly, Anthropic is expanding the range of tasks AI can handle — from sending files and pulling reports to checking builds and testing app flows.

The broader signal is hard to miss: Anthropic is steadily moving Claude toward becoming not just an assistant that answers questions, but an agent that can operate software on a user’s behalf.