Claude Cowork and Claude Code just took a very literal step toward becoming actual coworkers: they can now sit “at” your Mac, move the mouse, type into apps, open your browser, and quietly get work done while you do something else. It’s the closest the current crop of AI tools has come to feeling like a real digital assistant that doesn’t just talk about tasks, but actually goes and clicks the buttons for you.

Instead of wiring up a dozen integrations or Zapier-style automations, you can now give Claude permission to control your computer directly from within Claude Cowork and Claude Code. When it doesn’t have a native connector for something — say a niche analytics dashboard or a finicky internal tool — Claude will simply treat your screen like you do: it scrolls, points, clicks, opens files, and navigates across apps to complete whatever you ask. It can pull up your browser, run dev tools, or dig through folders with no additional setup beyond turning the feature on.

This is all rolling out as a research preview for paying users, so it’s very much “early access with training wheels.” The feature is available to Claude Pro and Claude Max subscribers, and at least for now, you need to be on macOS with the desktop app awake and running in the background. That Mac essentially becomes a host machine: once it’s online, you can hand Claude tasks from your phone and let it work on your desktop while you’re elsewhere.

A big part of this story is Dispatch, a feature in Claude Cowork and now Claude Code that lets you keep one continuous conversation going with Claude across devices. Picture this: you’re on the train, you message Claude from your phone to “pull my weekly metrics from the analytics dashboard and drop them into the report doc,” then you close your phone and move on with your morning. By the time you sit down at your desk, Claude has already opened your browser on your Mac, navigated to the right tools, fetched what it needs, and left the finished work waiting for you.

Dispatch becomes even more interesting once you think beyond basic automation. You could have Claude routinely check your email every morning, prep a briefing, spin up a Cowork session to outline a report, or kick off a Claude Code task to tweak a pull request. For developers, that might mean telling Claude to open your IDE, make a series of code changes, run tests, and prepare a PR — all initiated from your phone, executed on your Mac, and reviewed by you later. For makers, it could be something like managing a 3D printing project in stages without babysitting every step at your desk.

Of course, “AI that controls your computer” triggers every security instinct, and Anthropic is clearly trying to get ahead of that. The system is designed so Claude always asks for explicit permission before accessing a new app, and you can stop it at any time if something looks off. On the back end, Anthropic says it’s scanning model activations to detect things like prompt injection attempts, adding an extra layer of defense on top of those permission prompts.

Even with those guardrails, Anthropic is blunt that computer use is much earlier in its lifecycle than Claude’s text and coding abilities. It can and will make mistakes, it may bumble through more complex workflows, and controlling the screen is naturally slower and more brittle than hitting a polished API. The company recommends starting with apps you trust, avoiding sensitive data, and being aware that some categories of apps are blocked outright by default for safety reasons.

What’s interesting is how this shifts the mental model for AI in everyday work. Until now, most people have experienced tools like Claude as smart chat boxes, maybe with some plugin-style integrations for specific services. With computer control in Cowork and Claude Code, Claude becomes more like a remote teammate that can sit at a machine and grind through the boring operational steps: filling dashboards, copying data between windows, kicking off builds, or managing repetitive admin tasks. It’s not perfect, and it won’t magically run your whole job yet, but it pushes AI one layer deeper into the practical, messy parts of computer work instead of keeping it confined to the chat window.

If you’re already using Claude Pro or Max on a Mac, the on-ramp is intentionally low friction: update your desktop app, make sure it’s running, pair it with the mobile app, and start by offloading one or two small but annoying tasks to see where it shines and where it stumbles. You’ll quickly spot its sweet spots — structured workflows, repeatable browser actions, dev tool routines — and its rough edges where it still feels like a clumsy intern. But as a glimpse of where AI assistants are headed, Claude quietly moving your mouse in the background might be one of the more important and tangible steps forward.