Claude can now control your computer while you assign tasks from your phone

Imagine assigning a coding task from your phone while riding the train, and arriving at your desk to find the tests passing and the pull request ready for review.

Anthropic is turning that workflow into reality. The company just rolled out its computer use capability in research preview for Claude Cowork and Claude Code, available to Claude Pro and Claude Max subscribers.

Pointing, clicking, and thinking like a human

Instead of requiring complicated setup or bespoke integrations, Claude can now interact directly with your screen. It moves the mouse, clicks buttons, scrolls through feeds, opens local files, and fires up developer tools on its own.

Claude tries to be smart about efficiency. Whenever possible, it reaches for dedicated service connectors first, such as integrations with Slack or Google Calendar.

Direct screen control is reserved as a fallback when no API exists. Anthropic openly admits that navigating a visual interface is naturally slower than API calls, and complex requests might take a couple of attempts to get right.

Safety is obviously the elephant in the room when an AI takes over your mouse. Anthropic built in safeguards that monitor internal model activations to catch potential prompt injections. Claude is also programmed to ask for explicit permission before touching new apps, keeps certain sensitive tools off-limits by default, and can be stopped at any moment.

The phone-to-desktop handoff

While raw computer use was first demonstrated back in 2024, the real unlock here is how it ties into Dispatch, Anthropic's recently launched cross-device feature.

Dispatch lets you maintain a single running thread between your mobile phone and your computer. You can fire off a quick prompt on your phone to pull weekly metrics, organize email briefings, or check on a 3D printing project, while Claude does the actual heavy lifting on your desktop.

To make it work, your computer needs to stay awake with the desktop app running. While Anthropic notes broad desktop support in its post, The Verge reported that the preview is currently limited to macOS.

Why it matters

Anthropic is solving one of the biggest bottlenecks facing autonomous agents: the messy, API-less real world.

Most enterprise software and legacy tools never bothered to build modern integrations for AI. By teaching Claude to literally look at pixels and click buttons, Anthropic bypasses the integration hurdle entirely, making automation accessible without technical plumbing.

Combined with Dispatch, it also redefines what prompt engineering feels like in daily work. Instead of sitting in front of your terminal watching tokens generate, you can genuinely delegate tasks asynchronously from your pocket.

That said, screen-scraping autonomy comes with clear caveats. It is inherently slower, prone to breaking on unexpected UI updates, and carries security risks despite new model safeguards. For now, it is an impressive preview worth testing on low-stakes administrative work, but you probably should not hand it the keys to production just yet.