Claude Can Now Control Your Mac While You Sleep

I sent a single message from my iPhone while waiting for a subway, and 90 seconds later my MacBook Pro — sitting on my desk at home with the lid open — had launched Numbers, populated 30 cells with formatted revenue data, applied conditional formatting, and saved the file to iCloud. Nobody touched the keyboard. Nobody was in the room. That's Claude Computer Use, and it shipped on March 24, 2026.

The Gap Between Chat and Action

For the past three years, AI assistants have been extraordinarily good at one thing: generating text in a chat window. Ask a question, get an answer. Request a draft, receive prose. But there has always been a hard boundary between what AI could say and what AI could do. Computer Use erases that boundary.

If you've been following Anthropic's recent releases, you'll know about Dispatch — the feature that lets you message Claude from your phone and assign tasks to your Mac.

Post not found or has been removed.

Computer Use is what happens after Claude receives that message. It doesn't just think about your request and type a response. It physically operates your Mac. It opens applications. It clicks buttons. It types into text fields. It navigates browser tabs. It drags files. The operative word here is "operates" — not "suggests" or "recommends," but actually does the thing.

How the Screenshot Loop Works

The mechanism underneath is elegant in its simplicity. Claude captures a screenshot of your Mac's display, runs vision analysis to understand what's on screen, decides on the next action, then executes that action through system-level input events. The entire cycle takes a few hundred milliseconds, and it repeats continuously until the task is complete.

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│  Screenshot │ ──→ │  Vision     │ ──→ │  Action     │
│  Capture    │     │  Analysis   │     │  Execute    │
└─────────────┘     └─────────────┘     └──────┬──────┘
       ↑                                        │
       └────────────────────────────────────────┘
                     Loop

This is structurally identical to how a human uses a computer. You look at the screen, you decide what to do, you move the mouse and type. Claude does the same thing, except it processes visual information at pixel-level precision and supplements screen reading with macOS accessibility APIs. That means it can detect UI elements that are visually ambiguous or partially hidden — something even experienced human users occasionally struggle with.

The system runs inside Anthropic's Cowork environment, which provides the sandboxing and security layer.

Post not found or has been removed.

What You Can Actually Do With It

The Dispatch-to-Computer-Use pipeline unlocks workflows that previously required you to be physically at your desk. Here are the scenarios I've found most compelling after a day of testing.

Spreadsheet automation is the killer feature right now. I messaged Claude from my phone: "Open the CSV file in Downloads called sales-q1-2026.csv in Numbers, sort by column C descending, add a SUM formula in cell D32, and generate a bar chart of the top 10 rows." Two minutes later, it was done. The same task takes me about 25 minutes when I do it manually, between figuring out the sort, getting the formula right, and fighting with chart formatting.

Browser-based research works surprisingly well. I asked Claude to find pricing information for three competing SaaS tools, and it opened Safari, navigated to each site, located the pricing pages, and compiled the data into a note in the Notes app. The entire process took about four minutes. It handled cookie banners, navigation menus, and even a site that required scrolling to reveal pricing tiers.

Meeting preparation is another strong use case. A message like "check my calendar for today's meetings and create a Keynote outline for the 2pm product review" triggers Claude to open Calendar, read the event details, then switch to Keynote and build a slide deck with relevant section headers.

The Permission Model That Makes It Viable

Security is where Computer Use either succeeds or fails, and Anthropic made a deliberate choice here. Every time Claude tries to access an application it hasn't used before, it sends a permission request to your phone. You see a notification, you approve or deny, and only then does Claude proceed.

┌───────────────────────────────────────────────┐
│  Before Computer Use     │  After              │
├───────────────────────────────────────────────┤
│  Human operates directly │  Claude operates    │
│  No per-app permissions  │  Per-app approval   │
│  No real-time audit log  │  Every action logged│
│  Mistakes are permanent  │  Pre-action confirm │
└───────────────────────────────────────────────┘

This introduces friction, and that friction is intentional. The practical move is to pre-approve your commonly used apps — Numbers, Safari, Calendar, Keynote, Notes — so that everyday workflows run uninterrupted. Keep the approval requirement active for apps that handle sensitive data: banking apps, password managers, anything with credentials.

Anthropic is explicit in their documentation: "Computer use is still early." They recommend starting with trusted applications and gradually expanding as you build confidence in the system's behavior. I think this is the right call. An overeager AI clicking around your banking app is nobody's idea of a good time.

Computer Use vs. OpenClaw

The competitive landscape here is straightforward. OpenAI's OpenClaw offers similar desktop-control capabilities, but the architectural approach differs. Claude Computer Use is built around the mobile-first Dispatch workflow — you trigger tasks from your phone, they execute on your Mac. OpenClaw is more desktop-native, designed for automating tasks while you're sitting at the computer.

Neither approach is strictly better. If your workflow involves a lot of "I'm away from my desk but need something done on my computer," Claude's Dispatch integration is genuinely useful. If you're at your desk and want AI to handle tedious subtasks in parallel while you focus on something else, OpenClaw's model might fit better. My prediction is that both companies will converge toward supporting both use cases within the next two quarters.

What Breaks

Honesty time. Computer Use is not production-grade software for mission-critical workflows. Complex multi-step tasks occasionally go sideways — Claude might click the wrong menu item, misidentify a button, or get confused by an unexpected dialog box. Display resolution changes between your external monitor and laptop screen can throw off coordinate mapping. Dark mode to light mode transitions sometimes degrade recognition accuracy.

The macOS-only limitation is the most significant barrier to adoption. Developers on Linux and enterprise users on Windows are entirely shut out for now. Anthropic hasn't announced a timeline for cross-platform support, and given the deep integration with macOS accessibility APIs, I wouldn't expect a quick port.

You also need a persistent internet connection on both ends — your Mac and your phone. There's no offline queue, no delayed execution. If your Mac loses connectivity, in-progress tasks will fail.

Getting Started Without Getting Burned

Start with trivial tasks. Ask Claude to open a specific file, copy a paragraph of text, or perform a simple web search. Watch how it navigates your particular Mac setup — your dock layout, your menu bar apps, your default browser. Once you've confirmed the basics work reliably, graduate to multi-step workflows.

The single most important thing you can do is write specific instructions. "Clean up my files" is a recipe for confusion. "Move all PDF files from Downloads to Documents/Invoices and rename them with today's date as a prefix" is a recipe for results. Think of Claude as a capable but brand-new hire who needs explicit instructions until they learn your preferences.

Why This Matters Beyond Productivity

Computer Use represents a category shift in what AI can do. For three years, AI has been a text-in, text-out system. Brilliant at language, helpless at action. Computer Use breaks that constraint. An AI that can operate your computer is not a chatbot — it's an agent in the original, meaningful sense of the word. It acts on your behalf in the digital world.

Anthropic's strategy is now fully visible. Cowork provides the environment. Dispatch provides the communication channel. Computer Use provides the hands. Together, they form something that looks less like a chat assistant and more like a digital employee who happens to work for free, never sleeps, and already has access to your Mac.