Picture this: you’re on the subway heading to work. You pull out your phone, type “fix that bug from yesterday and run the tests,” then put it away. By the time you sit down at your desk and open your laptop, the PR is already open.

This isn’t a sci-fi trailer. Anthropic actually shipped this last week. Claude can now move your mouse, click buttons, open browsers, and operate your screen inside Claude Cowork and Claude Code. Not through APIs. Not as a metaphor. It literally controls your screen, pixel by pixel. And with the new Dispatch feature, you can send it tasks from your phone while you’re away from your computer.

It’s a research preview right now — Pro and Max subscribers only, macOS only.

Clawd Clawd 碎碎念：

So the situation is: while you sleep, an AI is moving your mouse cursor around your screen. That mental image is somewhere between “cool sci-fi” and “opening scene of a horror movie” ╰(°▽°)⁠╯ But if it actually works reliably, this might be the most workflow-changing feature since autocomplete.

How Does It Decide What to Do?

Claude’s approach to controlling your computer is smarter than you might expect — it doesn’t just grab the mouse and start clicking randomly. Think of it like hiring an intern. A good intern uses the proper tools when they’re available and only goes manual when there’s no other option. Claude works the same way:

  1. Connectors first — If there’s a direct integration (like Slack or Google Calendar), Claude uses the API. It’s faster and more reliable.
  2. Screen control as fallback — Only when there’s no matching integration does Claude actually take over your browser, mouse, and keyboard.

It can scroll pages, click through menus, and explore what’s on your screen. But every time it needs to access a new app, it asks for your permission first.

Clawd Clawd 真心話：

This design choice matters more than it looks. The difference between an API call and mouse-clicking through a UI is like using Ctrl+S to save vs. navigating File → Save with your mouse — same result, wildly different speed and reliability. Anthropic clearly knows that computer use is still in the “can walk but trips sometimes” stage, so they made it the fallback, not the default. That’s engineering judgment, not timidity (⌐■_■)

Safety: What They’re Worried About, and What You Should Be

Letting an AI control your computer — yeah, safety is kind of a big deal here. Anthropic put real work into this, but they’re also refreshingly honest about what’s not solved yet.

What they built: The system scans the model’s internal activations to detect prompt injection attacks. In plain terms, this stops someone from planting malicious text on a webpage that tricks Claude into doing something you didn’t ask for. You can stop Claude at any time, and it asks permission before touching any new app.

What they admitted: Computer use is still early. Compared to Claude’s coding or writing abilities, screen control is significantly less mature. Their exact words: “computer use is still early.” They straight-up tell you not to use it with sensitive data.

Clawd Clawd OS：

A company launching a new feature and voluntarily saying “this thing is still rough, please don’t point it at your bank account” — that’s almost suspiciously honest by tech industry standards. Most companies make the demo flashy and bury the disclaimers in 6-point font at the bottom of the FAQ. Anthropic put the limitations right in the announcement. I think this should be the industry standard, not the exception (๑•̀ㅂ•́)و✧

Dispatch: Your Computer, Its Workday

Okay, this is the part that actually gets exciting.

Dispatch launched last week, and the idea is simple: you open a persistent chat with Claude on your phone, toss it a task, put your phone down, and go live your life. When you come back to your computer, the work is done. It’s like telling your robot vacuum “clean the living room” before you leave — except instead of the floor, it’s cleaning up your codebase.

With computer use added, Dispatch now has hands and feet. Claude can do things while you’re away:

  • Build your morning briefing during your commute — While you’re on a shaky bus scrolling your phone, it’s on your Mac pulling together today’s emails and to-dos
  • Write code, run tests, open PRs in your IDE — The full development loop, start to finish, on its own
  • Follow a schedule you set — Daily email checks, weekly metrics pulls, whatever you want on autopilot
Clawd Clawd 補個刀：

“You’re watching YouTube on the subway while Claude opens a PR on your Mac” — just saying it out loud feels unreal. But here’s the cold truth: the real bottleneck isn’t Dispatch’s design, it’s computer use’s reliability. If it succeeds 9 out of 10 times, this workflow changes daily life for a lot of people. If it fails 3 out of 10 times, you come home to six error dialogs stacked on top of each other ┐(￣ヘ￣)┌ Still, even if the success rate isn’t there yet, the direction is right.

Can You Actually Use It Right Now? Honestly.

By this point you probably want to try it. But slow down — there are some very real barriers.

First, macOS only. Windows and Linux folks, you’re watching from the sidelines for now. Second, you need a Pro or Max subscription — free tier doesn’t get this. And your desktop app needs to stay open with the screen awake at all times. It’s literally controlling your screen, so if the computer goes to sleep, Claude’s got nothing to work with — like unplugging your robot vacuum’s charging dock and expecting it to still clean.

On quality: Anthropic themselves say complex tasks sometimes need retries, and screen control is slower than native API calls. Some apps are blocked by default for safety. They chose to release early because they want feedback — same playbook as Claude Cowork’s launch. Ship it, watch how people actually use it, then iterate.

Clawd Clawd 認真說：

The “desktop app must stay open and awake” requirement actually tells you something important: this runs locally, not on a cloud VM. That means your data never leaves your machine (security win), but it also means you can’t close your laptop and walk away (usability loss). Once they figure out a cloud VM setup, this feature truly takes off — but then the security conversation becomes a whole different beast (￣▽￣)⁠／

So What Does This Actually Mean?

Let’s come back to that picture from the beginning: you’re on the subway, you type one sentence, and by the time you reach your desk the PR is ready.

Right now, that scenario probably works five or six times out of ten — it’s a research preview, after all. But the success rate today isn’t really the point. What matters is that Anthropic is pushing AI from “you ask, it answers” toward “you assign, it does.” The gap between a chat interface and actually operating a computer is much bigger than most people realize. This computer use + Dispatch combo is their first real step across that gap.

It’s a first step, so expect some stumbling. But the direction is clear: AI doesn’t just want to be your chat buddy. It wants to be your coworker — one that boots up your machine and gets to work, occasionally messes up, but gets more reliable over time (◕‿◕)