How to automate your desktop with Claude's computer use feature

Picture this: you type a sentence to Claude, look away for a minute, and come back to find it actually clicked through a spreadsheet, opened your Notes app, and wrote a report — without you touching the mouse once. That's not a hypothetical anymore. It's what happens when you turn on Claude's computer use feature and let it drive your screen the way a person would.

This guide walks through what computer use actually is, how to set it up correctly, real tasks worth automating, and the privacy tradeoffs you need to understand before you hand Claude the keyboard.

What is Claude's computer use feature, exactly?

Computer use is Anthropic's capability that lets Claude see your screen and control your mouse and keyboard directly. Instead of you copying data between a chat window and your apps, Claude can take control of your Mac, clicking buttons, typing text, opening apps, and completing tasks on your screen while you tell it what you need done, and it does the work directly on your desktop. It launched as a research preview alongside a related feature called Dispatch, on March 23, 2026 as a research preview for Pro and Max subscribers, announced alongside Dispatch, Anthropic's new phone-to-desktop task delegation system.

Officially, computer use lives inside two surfaces: Claude Cowork and Claude Code, both accessed through the Claude Desktop app. The Anthropic computer use tool documentation describes it plainly: give Claude screenshot, mouse, and keyboard control of a desktop environment with the computer use tool, which provides screenshot capabilities and mouse/keyboard control for autonomous desktop interaction.

It's worth being honest about what this is not. It's not magic. Anthropic frames it as a fallback, not a first resort — in Cowork, Claude uses the most precise tool first: connectors if available, then the browser, then direct screen interaction, so computer use matters, but it is clearly not meant to be the first tool for every job.

How does Claude actually see and control your screen?

Claude doesn't have some special hook into your operating system. It works the same way a remote human assistant would: it looks, then it acts. To control your Mac, Claude constantly takes screenshots to see what is on your monitor, which means the AI can see your private files, photos, or messages if they are open on your screen.

Once it has a screenshot, it decides what to click, type, or scroll. On the technical side, the computer use tool reference notes that enhanced actions available in newer tool versions include scroll with amount control, click-and-drag between coordinates, and additional mouse buttons like right-click and middle-click. That's a meaningfully richer toolkit than the original 2024 version, which was limited to basic clicks and keystrokes.

Anthropic also built in a priority order so Claude isn't blindly clicking around when a cleaner option exists. Claude uses a smart priority system: first it tries direct integrations like Slack or Gmail, then it uses the web browser to navigate, and finally, as a last resort, it looks at the screen and clicks or types — a layered approach that makes it more reliable and efficient.

What do you need before you start?

Before enabling anything, make sure you actually qualify. This feature currently has real restrictions:

A paid plan. A Pro plan ($20/month) or Max plan (from $100/month) is required — Computer Use is available on both personal and team plans.
The right platform. Computer use in Claude Code specifically is a research preview on macOS that requires a Pro or Max plan, is not available on Team or Enterprise plans, and requires Claude Code v2.1.85 or later in an interactive session. Desktop app support for Windows arrived later in 2026, but macOS remains the most mature experience.
The Claude Desktop app, downloadable from claude.ai/download.
macOS permissions. Claude needs Accessibility (to click and type) and Screen Recording (to see your display) permissions granted through System Settings.

How do you set up Claude computer use step by step?

Here's the actual sequence, based on Anthropic's own product flow and independent setup walkthroughs:

Open the Claude Desktop app and make sure it's updated.
Click your profile icon in the bottom corner and open Settings.
Find the Computer use toggle and turn it on. You'll get a confirmation dialog — a confirmation dialog explaining what Claude will be able to do, including taking screenshots and controlling your keyboard and mouse, appears, and you click "Turn on" after reading it.
Grant the Accessibility and Screen Recording permissions when macOS prompts you. If you're using Claude Code from the terminal, the Claude Code computer use docs note that the first time Claude tries to use your computer, you'll see a prompt to grant two macOS permissions — Accessibility and Screen Recording — and macOS may require you to restart Claude Code after granting Screen Recording.
Before you run anything sensitive, check the Denied apps list in settings. Look at the Denied apps section and add any app you don't want Claude to interact with — at minimum, add your password manager.
Open a Cowork session (or Claude Code, for developer-focused tasks) and type a plain-English instruction, like organizing files or filling out a form.
Approve the first action. Claude won't barge into new apps unannounced — Claude operates safely and will always ask for your explicit permission before opening a new app on your machine for the first time.

What can you actually automate with it?

This is where computer use gets genuinely useful instead of just impressive. A few realistic categories:

Local file and note tasks. One documented test had Claude open the Notes app, create a note with a specific title, and fill it with exact text — Claude first decomposes the task, shows a one-line status, and waits for approval on the specific command before any keystroke or mouse event reaches another app. The whole loop, from prompt to verified result, took under ten seconds, with the only human action being the "Allow once" click.

Browser research and dashboard checks. You can ask Claude to pull numbers from a web dashboard without writing any code. Click Cowork to enter the mode where Claude can work with your files, browse the web, and control your computer — you can type something like "Go to my Google Analytics in Chrome and tell me how many visitors I had yesterday" and Claude will navigate there, read the dashboard, and report back.

GUI app testing and debugging (developer use case). This is arguably the strongest fit right now. Through the CLI, you can ask Claude to build a macOS menu bar app — it writes the Swift, compiles it, launches it, and clicks through every control to verify it works before you ever open it. It's also handy for visual bugs: tell Claude the modal is clipping on small windows, and it resizes the window, reproduces the bug, screenshots it, patches the CSS, and verifies the fix — Claude sees what you see.

Remote task delegation via Dispatch. If you want to kick off a desktop task from your phone, Dispatch is the companion feature. Setup involves toggling on file access, keep-awake, Chrome access, and computer control, then scanning a QR code with your phone to sign into the same account, after which you can type a message to Claude from your phone and it will get to work on your computer, showing progress in real time on both devices.

How does Dispatch let you run tasks from your phone?

Dispatch essentially separates "telling Claude what to do" from "being at your desk." Your computer needs to stay awake for Dispatch to work — the setup enables a "Keep awake" toggle that prevents your Mac from sleeping while Dispatch is active. That's a real tradeoff: leaving a machine awake and unattended all day has its own security implications, especially in shared workspaces.

What are the safety and privacy risks?

This is the part that deserves more attention than most tutorials give it. Because Claude works from screenshots, it can see everything visible on your screen, not just the app it's working in. Anthropic strongly warns against using this feature for any financial, medical, or legal tasks because small mistakes can cause big problems, so you should always close sensitive documents before giving Claude a task.

There's also the prompt injection risk — malicious instructions hidden inside a webpage or document that trick the model into taking unintended actions. Anthropic has built defenses for this directly into the model. According to the official documentation, Anthropic has trained the model to resist prompt injections and added an extra layer of defense — classifiers automatically run on prompts to flag potential injections, and when they identify one in a screenshot, they steer the model to ask for user confirmation before proceeding. Even so, these precautions remain important even with the classifier defense layer in place, and you should inform end users of relevant risks and obtain consent before enabling computer use in your own products.

On the data side, if you're worried about Anthropic storing your screen activity, note that computer use is a client-side tool — all screenshots, mouse actions, keyboard inputs, and any files involved in a session are captured and stored in your environment, not by Anthropic. For enterprise users, the feature is also eligible for Zero Data Retention, meaning data sent through it is not stored after the API response returns when an organization has that arrangement.

How does Claude computer use compare to other AI automation tools?

Claude isn't the only player here, and the comparison actually clarifies what each tool is for rather than which one "wins."

ChatGPT Agent (formerly Operator) is browser-only but scores higher on pure web navigation benchmarks. Claude Computer Use trails both on browser-only tasks but is the only one of the three that can drive an entire desktop, not just a browser tab. On the older head-to-head numbers, Operator achieved a leading 87% on browser automation while Claude Computer Use scored 56%, indicating moderate performance in browser-specific tasks.
Gemini Computer Use leans hard into browser workflows. Google's Gemini Computer Use, grown from Project Mariner, optimizes for browser workflows where DOM awareness and web-native actions outperform generic screen scraping.
Browser Use (open-source) is the cheapest option if you only need web automation. If you only need to automate websites and want control and low cost, Browser Use is the natural starting point, and it's free to self-host — you pay only the API cost of whatever model you point it at, making it almost always the cheapest path for high-volume browser automation.

The honest bottom line: Claude's edge is that it isn't locked to a browser tab. If you need to operate desktop applications beyond the browser, only Claude Computer Use can do it. Reliability has improved dramatically since launch, too — the original Claude 3.5 Sonnet scored 14.9% on OSWorld in late 2024, while Anthropic's recent Opus models reach the low 80s, with Opus 4.7 reported at 82.3%, though that still means roughly one in five desktop tasks fails.

What are the current limitations you should plan around?

Even with the benchmark gains, don't build critical, unattended workflows on this yet. A few structural limits worth knowing:

Availability restrictions. The feature is not available for Free plan users, Team accounts, or Enterprise accounts as of its research preview status.
Reliability variance. Complex tasks sometimes need a second try — computer use is powerful, but still operationally fragile enough that you should not quietly build critical unattended workflows on top of it without testing hard.
Uptime requirements. Your desktop must be active — Anthropic says your computer needs to be awake and the Claude Desktop app needs to be open.
Preview instability. The preview status means availability and behavior can change quickly, so if you're writing process around it for a team, treat it as evolving, not frozen.

What best practices should you actually follow?

A short checklist worth pinning somewhere before your first real session:

Close anything sensitive — banking tabs, medical portals, personal messages — before starting.
Add password managers and financial apps to the denylist immediately.
Start with small, verifiable tasks (a note, a form, a single dashboard check) before trusting it with anything multi-step.
Stay in the room, at least for now. For best results, close sensitive apps before starting a computer use session, start with simple tasks to see how Claude handles them, and always review Claude's actions as they happen — you can stop Claude at any point.
Save prompts that work well. When you find a prompt that works well, save it as a scheduled task or keep it somewhere handy.

Claude's computer use feature is the clearest sign yet that AI assistants are moving from "tell me what to type" to "watch me do the work myself." It's not flawless, and it shouldn't be trusted with your bank account yet — but for the boring, repetitive click-through tasks that eat up your afternoon, it's already good enough to hand off. The real skill now isn't prompting Claude to write something for you. It's learning to write instructions clear enough that an AI can act on them without supervision — and that's a habit worth building before everyone else catches up.