Claude Just Got Access to Your Computer. Here's What That Actually Means

Anthropic announced this week that Claude can now control your computer and complete tasks on your behalf. Open apps, browse the web, fill out spreadsheets, attach files to calendar invites. You send it a task from your phone and come back to finished work on your desktop.

The demo they showed was someone running late for a meeting who asked Claude to export a pitch deck as a PDF and attach it to the invite. Claude did it. No one touched the keyboard.

This is the thing AI companies have been promising for years and largely failing to deliver. Whether Anthropic has actually cracked it or just made a slick demo video is a different question, and one that won't get answered until regular people start using it and reporting back. Early access products from AI companies have a history of looking great in controlled conditions and falling apart when actual users get their hands on them. That pattern is worth keeping in mind here.

What the Feature Actually Does

The feature is available right now as a research preview for Claude Pro and Max subscribers on Mac. Windows support is coming later. It works through Claude Cowork and pairs with a mobile tool called Dispatch that lets you assign tasks from your phone.

When Claude gets a task it first looks for a native integration like Google Calendar or Slack. If one exists it uses it. If not, it does what you would do: moves the cursor, clicks, types, navigates the screen manually. It reads the interface the same way you would and figures it out from there. Anthropic describes this fallback behavior as the system using the screen to navigate the same way a person would, which sounds straightforward but is genuinely hard to pull off reliably across different apps, operating systems, and screen layouts.

For developers specifically, Claude can work inside an IDE, submit pull requests, and run tests. The idea is that you hand off the mechanical parts of a workflow and stay focused on the decisions that actually require your judgment.

Where This Came From

Anthropic acquired a startup called Vercept in February 2026. Vercept had already built a cloud-based agent capable of operating a MacBook remotely, which is almost exactly what this feature is. The computer use announcement shipped less than four weeks after that acquisition closed. One of Vercept's co-founders posted on X that the team had only been at Anthropic for four weeks when they shipped it.

That timeline matters because it tells you this wasn't a slow internal research project that finally made it to production. Anthropic saw what Vercept built, bought it, and moved fast. That's a different kind of signal than a feature that spent two years in development. It means they identified a gap, found someone who had already solved it, and prioritized getting it out the door.

Why Everyone Is Rushing to Do This Right Now

A tool called OpenClaw went viral earlier this year. It lets people assign tasks to AI through WhatsApp and Telegram and it runs locally on your device, giving it access to your files and apps. Nvidia's CEO called it the next ChatGPT. The response from the industry was immediate. Every major AI company started accelerating whatever agentic work they had in progress.

Anthropic's computer use launch is their answer to that moment. The underlying message is that Claude can do what OpenClaw does and more, without requiring you to route everything through a third party messaging app. Whether that positioning holds up is something the market will sort out over the next few months.

The Safety Stuff

Claude is required to ask permission before accessing any new application. It shows you the plan before it starts executing. You can stop it at any point. Anthropic has also built in defenses against prompt injection attacks, which is the scenario where Claude is browsing a webpage or reading a document and the content of that page tries to trick it into doing something it shouldn't. That's a real attack vector for any AI that can take actions on your behalf, and it's good that they're thinking about it explicitly rather than treating it as an edge case.

They're also being honest that this is not a finished product. Complex multi-step tasks will break. The system will make mistakes. Anthropic said as much publicly, which is either refreshing honesty or the kind of disclaimer you put out to manage expectations before the negative reviews start coming in. Probably both.

What Happens Next

The research preview is limited to Mac for now. Windows support is on the roadmap. The current rollout is intentionally narrow so Anthropic can collect feedback before a broader release, which is the right call for something with this much access to a user's machine.

The longer term question is what this looks like when it works reliably. If an AI agent can genuinely handle the low-judgment repetitive computer work that eats hours out of a normal workday, that changes how a lot of people structure their time. Not in some abstract future sense. In a pretty practical near-term sense for anyone whose job involves a lot of routine file management, scheduling, data entry, or software tasks.

We're not at "hand it your login and go on vacation" territory yet. But the gap between where this is now and where it needs to be to be genuinely useful in daily work is smaller than it's ever been. Whether Anthropic closes that gap faster than the competition is the actual story, and that one's still being written.