Anthropic just shipped what Apple promised a year ago and still hasn't delivered: an AI that can actually control your Mac.

The Summary

  • Claude now has "computer use" capability in research preview for Pro and Max subscribers, letting it point, click, and navigate your screen to complete tasks without setup.
  • Works with Dispatch (assign tasks from your phone) and pairs with Claude Code and Claude Cowork, opening files, browsing, and running dev tools autonomously.
  • Available now. Apple's promised agent features are over a year late.

The Signal

This is the messy, dangerous, thrilling moment when AI agents stop being toys and start being risky enough to matter. Claude can now take the wheel on your Mac. Not through APIs or sandbox environments. Through your actual screen. It sees what you see, clicks what you'd click, opens what you'd open.

The technical leap here isn't huge. Screen control and computer vision models have been around. What's different is the packaging: Anthropic is betting that enough people will trust an AI to navigate their personal computing environment that making it easy, frictionless, and phone-assignable (via Dispatch) creates a new behavior. This is the agent economy's rubber-meets-road moment. Not "what if agents could do X someday" but "here, let Claude loose on your files and browser right now."

The timing matters. Apple announced agentic Siri features in 2024. We're in March 2026. Still nothing. Anthropic didn't wait for permission or polish. They shipped in research preview, knowing some percentage of users will absolutely break things or leak sensitive data or discover horrifying edge cases. That's the play: move fast, learn in public, and own the category before the slow-moving incumbent wakes up.

The Electron app complaint (from Daring Fireball's Gruber) is telling. Claude's Mac client is still garbage. But Anthropic isn't optimizing for Mac nerd approval. They're optimizing for "does this thing feel like magic to enough people that they forgive the rough edges." That worked for ChatGPT. It's working here. The gap between what you can do with Claude on Mac and what you can do with Siri is now a canyon.

The Implication

If you're building anything that requires repetitive computer work (data entry, research workflows, testing interfaces), pay attention. This isn't about replacing humans yet. It's about seeing where the line is between "too risky to automate" and "risky enough to try." Anthropic is drawing that line in public. Watch where users actually use this versus where they recoil. That map will tell you what gets automated next and what stays human for longer than you think.

Source: Daring Fireball