The short answer
Computer use is the ability of an AI model to operate a computer the way a person does: it looks at screenshots, then moves the pointer, clicks, types and opens apps to finish a task. In Claude, computer use is a beta on the Pro, Max and Ultra plans, switched on in Claude Desktop for macOS and Windows. Anthropic says Claude tries connectors first, then the browser, and uses the screen only when nothing more precise covers the task.
| Plain meaning | AI that sees your screen and uses the mouse and keyboard to do a task |
|---|---|
| Claude | Beta on Pro, Max and Ultra in Claude Desktop (macOS and Windows); not on Team or Enterprise yet |
| Where to switch it on | Claude Desktop: Settings > Capabilities (or Settings > System, under This computer) > Enable computer use |
| Cloud versions | ChatGPT dots and Perplexity Computer run tasks on a computer in the cloud |
| Main catch | No sandbox between Claude and your apps on your own machine, and prompt injection is an open risk |
What computer use means, in plain English
Most AI tools work through text: you ask, they answer, and you do the clicking. Computer use changes who does the clicking. The model takes a screenshot, works out where things are on the screen, and sends mouse and keyboard actions, then takes another screenshot to see what happened. It repeats this until the task is done or it needs your help.
That makes it possible for AI to work in software that has no integration at all: an old internal system, a desktop app, a website without an API. It also makes computer use the slowest and least precise way for an AI agent to act. Anthropic's own safety article says Claude uses "the most precise tool first": connectors (see MCP), then the browser, and only then the screen, because screen work is slower and more error-prone.
Where you can use computer use in November 2026
Claude on your own computer
In Claude, computer use is a beta for Pro, Max and Ultra subscribers. Anthropic's help article says Team and Enterprise plans do not have access at this time. To switch it on, update Claude Desktop, open Settings, then Capabilities (or Settings, then System, under "This computer") and turn on Enable computer use. It works on macOS and Windows, your computer must stay awake and the Claude Desktop app must stay open.
Claude asks before it accesses each app, and some apps, including investment, trading and cryptocurrency platforms, are blocked by default. On macOS 15 or later it can work in background windows, but it asks once per session before taking over the full screen. Since September 2026 this sits inside ordinary Claude conversations, where Claude decides whether to answer or take on a task. Our Claude guide covers the plans.
Agents with their own computer in the cloud
Other tools run the computer somewhere else, which keeps your own machine out of it. OpenAI's ChatGPT Work is an agent that works across your connected apps and files, and ChatGPT's dots are always-on agents with their own cloud computer for eligible Pro and Business Premium users aged 18 and over (Pro access excluded the EEA, Switzerland and the UK at launch). Perplexity Computer normally runs multi-step tasks in a secure cloud sandbox on paid plans. Since 13 October 2026, Google AI Pro and Ultra subscribers have a Tasks toggle in Gemini that keeps working across Google apps while you are away. See the ChatGPT, Perplexity and Gemini guides for plan details.
Why computer use matters at work
- It reaches software nothing else can. Copying figures from a legacy finance tool into a spreadsheet, filling the same web form for 30 suppliers or downloading a monthly report from a portal are classic jobs for it.
- It handles the boring middle of a task. You can ask for an outcome, such as "collect the last three invoices from each supplier portal into this folder", rather than writing out every click.
- It shifts your job to reviewing. You set the task, watch the first runs and check the result. That only pays off for tasks you repeat; for a one-off, doing it yourself is often faster.
The risks, in Anthropic's own words
Anthropic's computer use safety article is unusually direct. It says there is no sandbox between Claude and your apps, that Claude takes screenshots and can see anything on screen, that an action in one app can affect another (a link clicked in email may open in Chrome even without Chrome permission), that prompt injection is a continuing threat with imperfect safeguards, and that Claude can make mistakes and occasionally act outside its intended boundaries. Its advice: do not grant access to banking, healthcare or government apps; avoid tasks involving financial accounts, legal documents, medical information or other people's personal data; close sensitive files first; start simple; and watch Claude's work, especially early on.
The OWASP Top 10 for LLM applications names the underlying problem Excessive Agency (LLM06:2025): giving a model more permissions than the task needs. Grant one app at a time and only for the task in hand. Our guide to using AI at work safely turns this into a team checklist.
A safe first task you can try
If you have Claude Pro, Max or Ultra and Claude Desktop, try a task that touches nothing important.
- Close your email, banking and any confidential documents.
- Switch on computer use as described above.
- Create a folder on your desktop called "AI test" with three harmless files, such as a few public web articles saved as PDFs.
- Give Claude the prompt below and approve only the apps it needs, such as Finder or File Explorer and a text editor.
- Watch every step. Note where it was quick, where it hesitated and where a connector or a simple script would have done the job better.
If you are not on a paid Claude plan, you can still learn the pattern by watching how agents plan tasks in our guide to using AI agents and how to automate tasks with AI, which often solve the same job with no screen control at all.
Related terms worth knowing
- AI agent: AI that carries out multi-step tasks; computer use is one way agents act.
- MCP: a standard way to connect AI to apps, which is more precise than clicking.
- Prompt injection: hidden instructions on a page or in a file that try to hijack the agent.
- Multimodal AI: computer use relies on the model reading screenshots.
- Benchmark: how models are scored; Anthropic's Claude Opus 5.5 results include OSWorld 2.1.
- AI safety: the wider field that studies how to keep agents within bounds.
Common misconceptions about computer use
- "It is just a faster me." By Anthropic's own account, screen work is slower and more error-prone than connectors. Its value is that it keeps going while you do something else.
- "If it asks permission, it is safe." Per-app approval limits where Claude can go, but Anthropic says actions in one app can still affect others and prompt injection is not solved.
- "It is the best way to connect AI to my tools." Anthropic itself ranks it last, after connectors and the browser. If a connector exists, use it.
- "My company can switch it on for everyone." In Claude it is not available on Team or Enterprise plans at the time of writing, so check your plan before you plan a rollout.
- "Cloud agents carry no risk." A cloud computer protects your own machine, but the agent can still act in any account you connect to it. Grant the minimum.
Prompts to copy
A harmless first computer-use task
In the folder called AI test on my desktop, open each PDF, find its title and publication date, and create a new text file called index.txt listing them in a table with the file name. Do not open any other folder or app. Stop and ask me if anything is unclear.
You get: A safe, checkable task that shows how computer use behaves
Decide if a task suits computer use
Here is a task I do every [week]: [describe steps and apps]. Tell me whether it is better done with a connector or integration, a browser agent, screen-level computer use or a simple automation tool, and why. List the permissions each option would need and the riskiest step.
You get: A clear choice of method with the smallest set of permissions