Anthropic's Claude Now Operates Your Mac So You Don't Have To

Yesterday, Anthropic shipped a feature that closes the gap between "the AI knows what to do" and "the AI actually does it." Claude can now control your Mac. It opens apps, navigates browsers, fills out forms, and finishes tasks on your behalf. Even while you're away from your desk.

What Anthropic released

The feature is called Cowork. It's available in research preview for Pro and Max subscribers. You describe what you need. Claude interacts with your screen the same way you would.

In Anthropic's demo, a user running late for a meeting asked Claude to export a pitch deck as a PDF and attach it to a calendar invite. Claude opened the right apps, navigated the interface, and completed the task with zero human input.

It pairs with Dispatch, a mobile companion that lets you assign tasks from your phone. Text Claude what you need from the train. Your Mac handles the rest.

How it decides what to do

Claude prioritizes speed. It first checks whether a direct integration exists for the task, connectors like Google Calendar or Slack. API calls are faster and more reliable, so it uses those when available.

When no connector exists, it falls back to screen-based control. It reads your display, moves the cursor, types on the keyboard, clicks buttons, and navigates between applications. Same steps you'd take. Just automated.

Where it works and where it breaks

The capabilities are broad: opening files, browsing the web, running dev tools, submitting pull requests, interacting with most Mac apps. For developers and knowledge workers, that covers a significant chunk of daily busywork.

But the rough edges are real. Anthropic is transparent about them:

• Mac only. Windows and Linux support hasn't shipped yet.
• Screen control is slow. Direct API calls take milliseconds. Navigating a UI takes seconds. Complex workflows require patience.
• Mistakes happen. Multi-step tasks sometimes need a second attempt.
• Keep sensitive data away. Anthropic recommends sticking to trusted apps and avoiding confidential information during the preview.

A permission-first model adds a layer of safety. Claude asks before accessing any new application. You can stop it at any time.

The real shift happening here

Until now, AI assistants lived in chat windows. They could tell you what to do. They could draft an email. But they couldn't open your email client, paste the draft, and hit send. That boundary just disappeared.

Think about how much of your workday is pure screen choreography. Pulling data from one app into another. Reformatting documents. Updating calendar invites. Sending follow-ups. These tasks need attention, but they don't need your judgment. They're sequences of clicks, and sequences of clicks are exactly what agents excel at.

What to do with this

OpenAI, Google, and Anthropic are all building toward the same vision: AI that operates software autonomously. They're converging fast. The gap between "AI assistant" and "AI coworker" is shrinking with every release.

You don't need to hand over your entire workflow today. But you should start experimenting. Pick one repetitive task. Delegate it. See what works and what breaks. The teams that build this muscle now will move fastest when these tools graduate from preview to production.