Claude Can Now Use Your Computer: Computer Use Comes to Cowork and Claude Code

Anthropic shipped computer use to Claude Cowork and Claude Code today. It’s a research preview, macOS only, available to Pro and Max subscribers.

The short version: Claude can now see your screen, move the cursor, click buttons, type text, and navigate between applications. Open apps, fill spreadsheets, browse the web. The kind of tasks you’d normally do yourself, sitting at your desk.

This isn’t Claude’s first time controlling a computer. Anthropic first launched computer use as a public beta back in October 2024 with Claude 3.5 Sonnet. At the time it scored 14.9% on OSWorld, a benchmark that tests real-world computer tasks like navigating spreadsheets and completing multi-step web forms. It was impressive as a proof of concept but not something you’d trust with actual work.

That number is now 72.5% with Claude Sonnet 4.6, released in February 2026. The human baseline on OSWorld is roughly 72.4%. Claude went from fumbling with scroll bars to matching human performance in about 16 months.

How It Actually Works

Claude doesn’t stream your screen like a video call. It uses what Anthropic describes as a “flipbook” method: taking sequential screenshots and piecing them together to understand what’s happening on screen. To interact with elements, it counts pixels vertically and horizontally to figure out where to place the cursor before clicking.

Three tools work together behind the scenes: a computer tool for mouse and keyboard input, a text editor for file operations, and a bash tool for system commands. Claude can scroll, click, type, open applications, switch between windows, and navigate web pages.

There’s a priority system built in. Claude doesn’t jump straight to controlling your screen for every task. It tries connectors and integrations first (Google Workspace, Slack, Calendar), then falls back to browser interaction, and only resorts to direct screen control as a last resort. Direct integrations are faster and more reliable, so screen-level control is reserved for situations where no better option exists.

The Sandbox

This is the part worth paying attention to. Cowork runs inside a lightweight Linux virtual machine on your Mac, using Apple’s Virtualization Framework (VZVirtualMachine). Network ingress and egress are strictly controlled. Claude Code operates within this sandboxed environment.

The design is local-first. Your files stay on your machine. Nothing gets sent to Anthropic’s servers as part of the computer use workflow. Felix Rieseberg from Anthropic discussed the architecture on the Latent Space podcast, explaining how the VM provides isolation without requiring users to configure containers or security policies themselves.

Safety Controls

Anthropic built several layers of safety into this release:

• Permission per app: Claude asks before accessing each new application. You can say no at any step.
• Prompt injection detection: Automatic scanning of model activations to detect injection attempts. Sonnet 4.6 shows significant improvement in resistance to prompt injections compared to earlier models.
• Restricted apps: Certain applications are blocked by default and require explicit user approval.
• Destructive action pauses: Before performing destructive actions like deleting files, Claude pauses and sends a push notification asking for confirmation.
• Social media guardrails: Claude is steered away from posting to social media, registering domains, or interacting with government websites.

Anthropic classifies this feature at AI Safety Level 2 (ASL-2). They recommend not using it with sensitive information during the research preview, and they’re upfront that prompt injection defense remains “an active area of development in the industry.”

What It Can Do (and What It Can’t)

The capability list is broad: opening and editing files, navigating web browsers, filling out forms, building spreadsheets with working VLOOKUP and conditional formatting, managing multiple browser tabs, and running developer tools.

The limitations are real too. The flipbook screenshot approach means Claude can miss short-lived notifications or brief animations. MacStories’ early review described performance as “still 50/50” and “slow.” Dragging, zooming, and some advanced interactions remain difficult. The desktop app has to stay open and awake while Claude works. Sessions can’t be shared with others.

It’s a research preview, and it feels like one. Good enough to be useful for certain workflows, not yet reliable enough to hand the keys over and walk away.

Dispatch: Send Tasks From Your Phone

This pairs with Dispatch, announced on March 17. Dispatch creates a persistent conversation thread between the Claude mobile app and the desktop app. You can describe a task on your phone, and Claude picks it up on your Mac, using computer use to execute it.

The idea is that you’re away from your desk but need something done on your computer. Send it to Claude from your phone, and it handles the clicking and typing while you’re gone. Dispatch sends push notifications before performing anything destructive, so you can approve or reject from your phone.

The Vercept Acquisition

Worth noting: Anthropic acquired Vercept earlier this year, a company specializing in AI perception and software interaction. The co-founders (Kiana Ehsani, Luca Weihs, Ross Girshick) brought expertise in visual understanding and interaction, which directly supports the computer use capability. The jump from 14.9% to 72.5% on OSWorld didn’t happen by accident.

Context: The Benchmark Race

For reference, here’s how computer use benchmarks have progressed:

Date	Model	OSWorld Score
Oct 2024	Claude 3.5 Sonnet	14.9%
Feb 2026	Claude Sonnet 4.6	72.5%
Human baseline	—	~72.4%
Feb 2026	GPT-5.4	75.0%

GPT-5.4 is currently the first model to exceed the human baseline at 75%. Claude is right at parity. Both numbers were unthinkable 18 months ago.

Who Should Try This

If you’re on a Claude Pro or Max plan and you use macOS, you can enable this today in Claude Cowork or Claude Code. It’s a research preview, so set your expectations accordingly: useful for repetitive desktop tasks, form filling, and browser-based workflows where you’d otherwise be clicking through the same steps manually.

For anything involving sensitive data, credentials, or actions you can’t easily undo, wait for the feature to mature. Anthropic says as much themselves.

The official announcement has full details on getting started.