Anthropic announced on March 23 that Claude can now control your Mac — pointing, clicking, scrolling, and navigating your apps to complete tasks on your behalf. The feature is available in research preview for Claude Pro ($20/month) and Claude Max ($200/month) subscribers, and is currently macOS-only.
What It Does
Inside Claude Cowork and Claude Code, users can now grant Claude permission to operate their Mac directly. Claude can open and edit files, use the browser, run developer tools, take screenshots, fill in spreadsheets, and handle multi-step workflows — anything you’d do sitting at your desk.
Claude’s approach is layered. It starts by reaching for the most precise tool available: if there’s a direct connector to a service like Slack or Google Calendar, it uses that first. When no connector exists, it falls back to operating the screen itself — navigating the UI like a human would, with your mouse and keyboard.
Anthropic engineering lead Felix Rieseberg, who announced the release, highlighted the pairing with Dispatch, a feature that maintains a single continuous conversation between your phone and Mac. In a demo, a user sends a task from their iPhone — convert a pitch deck to PDF and attach it to a calendar invite — and returns to find it done. Claude handles the whole workflow on the desktop while the user is away.
Permission-First, With Caveats
Before accessing any app, Claude asks for explicit permission. Users can also stop it at any point. Anthropic has implemented automated checks to detect prompt injection attempts — a real concern with agent-style AI that can make broad, fast changes to a system.
That said, Anthropic is upfront about the limitations. “Computer use is still early compared to Claude’s ability to code or interact with text,” the company notes. “Claude can make mistakes.” The recommendation: start with apps you trust and avoid sensitive data for now. Certain apps that handle confidential information are disabled by default.
Security researchers have flagged that agent-style AI introduces real risk — an AI with mouse and keyboard access to your machine can reach files, browsers, and internal tools quickly. Anthropic says it will continue hardening safeguards as the preview evolves.
How It Compares
This puts Anthropic squarely in the agentic AI race. Perplexity’s Computer tool launched a similar capability last month, initially as a cloud-based agent and now with a local version requiring a dedicated Mac mini. The open-source OpenClaw agent has been viral for its ability to connect to messaging apps and execute tasks across platforms. Claude’s version is more locked down than OpenClaw — permission-gated, Mac-only for now — but it’s also far easier to set up, requiring no configuration.
Windows support is expected in the coming weeks. The feature builds on improvements to Claude Sonnet 4.6, which brought enhanced performance for computer-based tasks.
How to Try It
Download or update the Claude app for macOS from claude.ai/download. The computer use option appears in Claude Cowork and Claude Code once system permissions are granted. You’ll need an active Claude Pro or Max subscription.
