Cowork already lets Claude create files, run code, and work with connected services. But some tasks live in apps that don’t have connectors — your company’s internal dashboard, a design tool, a government portal, a desktop application without an API. Until now, those were out of reach.
Computer Use changes that. When enabled, Claude can see your screen, move the mouse, click buttons, type text, and navigate applications — the same way you would, but without you being at the keyboard. It’s still a research preview, with clear limitations, but for the right tasks it removes a real bottleneck.
How it works
Claude uses a three-tier approach to getting things done:
-
Connectors first. If a direct integration exists — Gmail, Google Drive, Slack, Asana — Claude uses it. This is the fastest and most reliable path.
-
Browser next. If no connector exists but the task can be done in a browser, Claude navigates the web version.
-
Screen control as fallback. Only when neither of the above works does Claude take screenshots of your screen, analyse what it sees, and interact through mouse clicks and keyboard input.
This hierarchy matters. Computer Use is powerful but slower and less reliable than a direct connector. Claude doesn’t resort to screen control when a better option is available.
Enabling it
Computer Use requires:
- macOS (Windows support is planned but not available yet)
- A Pro or Max subscription
- The latest Claude Desktop app
- Your computer must be awake with Claude Desktop running
To enable it: open Claude Desktop, go to Settings → General, and toggle Computer use on. That’s it. The next time Claude needs to interact with an app during a Cowork session, it will ask for permission before accessing it.
What Claude can do
Some concrete examples where Computer Use earns its keep:
Filling in web forms. A government portal where you need to submit the same type of form regularly. You describe what needs to be filled in, Claude navigates to the portal, fills in the fields, and pauses before submitting so you can review.
Working with desktop apps. Testing an app in a phone simulator, formatting a spreadsheet in a specific way that requires menu navigation, or interacting with a proprietary internal tool that only exists as a desktop application.
Multi-source research. Claude can open your browser, search multiple sites, cross-reference information with files on your desktop, and compile findings into a document — all without you switching between tabs.
Repetitive GUI tasks. Anything where you find yourself clicking through the same sequence of menus and dialogs repeatedly. Data entry across applications, exporting reports from a dashboard, updating records in a system that lacks a bulk-edit function.
Safety controls
Computer Use takes screenshots to navigate your screen. This means Claude can potentially see everything visible on your display — personal data, sensitive documents, open browser tabs, notifications. Anthropic has built in several safeguards:
Per-app permissions. The first time Claude wants to access an application, it asks for your explicit approval. You can deny access to any app.
Default blocklist. Investment platforms, cryptocurrency apps, and trading tools are blocked by default. You can add other apps to a custom blocklist.
Action review. You can stop Claude at any time. If something looks wrong, interrupt the session.
Training boundaries. Claude is trained to avoid stock trading, entering sensitive credentials, and interacting with financial or medical systems — though Anthropic notes these guardrails aren’t absolute.
What to watch out for
This is a research preview, and the documentation is clear about the limitations:
It’s slower than connectors. Screen interaction involves taking screenshots, analysing them, planning actions, and executing them one by one. A task that takes 2 seconds via a connector might take 30 seconds via screen control.
Complex tasks sometimes need a second try. If a page loads differently than expected, or a dialog pops up in an unusual location, Claude might get confused. Treat it like a capable but literal-minded assistant: if the first attempt fails, a second try usually works.
Don’t use it for sensitive apps. The documentation explicitly warns against using Computer Use with banking, healthcare, government, and legal applications. The risk isn’t that Claude will do something malicious — it’s that screenshots capture everything on screen, and mistakes in sensitive contexts have real consequences.
Your screen is visible to Claude. Close apps and tabs you don’t want Claude to see before starting a Computer Use session. If your email is open in the background with sensitive correspondence, Claude might inadvertently capture it in a screenshot.
When to use it (and when not to)
Good use cases: repetitive GUI tasks, form filling, testing apps, navigating tools without APIs, multi-source research across browser tabs and local files.
Not recommended: anything involving money (banking, trading), sensitive personal data (medical records, legal documents), or high-stakes actions where a mistake is expensive.
The sweet spot is tasks that are tedious for you but straightforward for a screen-reading agent: “go to this dashboard, export last month’s data, put it in a spreadsheet, and email it to me.” That’s exactly the kind of mechanical work that Computer Use handles well.
What’s next?
- New to Cowork? Start with the beginner’s guide
- Want to trigger Computer Use from your phone? Read about Dispatch
- Looking for tasks Claude can do without screen control? See 10 unexpected things