Anthropic has added computer control capabilities to its Claude AI assistant, allowing the software to click, type, scroll, and navigate applications on a user’s Mac on their behalf. The feature is available as a research preview for Claude Pro and Max subscribers.
Claude follows a priority system when completing tasks. It first uses direct connectors to services like Gmail, Slack, or Google Calendar. If no connector exists, it navigates through the Chrome browser. Only as a last resort does it interact directly with the screen.
The feature works alongside Dispatch, a recently released tool that lets users assign tasks to Claude from a mobile phone. Users pair their phone with their Mac by scanning a QR code. Claude then executes tasks on the desktop while the user is away and sends back results.
Anthropic has built in several safeguards:
- Claude asks for permission before accessing each application
- Users can stop Claude at any time
- Some sensitive apps are blocked by default
- The system scans for prompt injection attacks
However, Anthropic openly acknowledges limitations. Early testing by MacStories found the feature succeeded roughly half the time. The company recommends avoiding sensitive data and notes that screen-level interaction is slower and more error-prone than direct integrations.
The feature currently supports macOS only and requires the Claude desktop app to remain open and running.
Sources: Claude Blog, VentureBeat, CNet
