Products, laws and numbers here change fast. Checked October 2026.
In plain words
Most AI tools connect to other software through special doorways built for programs, called APIs. But most websites and apps were built for people, with buttons and forms, not doorways for programs.
Computer use lets an AI work through that human interface. The AI is shown a screenshot, decides what to do ("click the Guests box"), and sends a mouse or keyboard action. A new screenshot shows what changed, and it decides again. Step by step, it can fill in a form, compare prices across sites or move data between apps.
It is slower and more error-prone than a direct connection, and it needs care: an AI that can click anything can also click the wrong thing. Products that offer it ask you to confirm purchases and take over for passwords and payments.
Everyday example
Helping a relative over a video call
Your uncle shares his screen and asks you to book his train ticket. You can't touch his laptop, so you look and say: "Click the blue Search button. Now type 4 in passengers." You watch the screen change after each step and guide the next. Computer use is the AI doing both parts: looking, then moving the mouse itself.
Where the example breaks: You would recognise a fake pop-up or a scam page instantly. An AI may treat text on a malicious page as instructions, which is why providers advise running it in a separate, locked-down environment.
How it actually works
Computer use is a form of tool use in which the tool is a whole screen, run in a loop like any agent.
See
The app takes a screenshot of a browser or desktop, usually inside a virtual machine, and sends it to a multimodal model that can read images.
Decide
The model works out where it is in the task and picks one or a few actions, such as click at a position, type some text, scroll or press a key. Positions are given as pixel coordinates on the screenshot.
Act
The app carries out the actions on the real machine, then takes a new screenshot. The loop continues until the task is finished, a step limit is hit, or the model needs you.
Stay safe
Anthropic's documentation advises a dedicated virtual machine with minimal access, keeping login details away from the model, limiting which sites it can visit, and keeping a human in the loop for consequential actions. Its main warning is prompt injection: instructions "on webpages or contained in images" might override yours.
- Launched
- Anthropic released computer use in public beta for developers on 22 October 2024, with Claude 3.5 Sonnet
- Early scores
- On the OSWorld test of computer tasks, that first version completed 14.9% using screenshots only
- OpenAI
- Operator launched in January 2025, then folded into ChatGPT agent in July 2025
- Released a Gemini 2.5 Computer Use model for developers in October 2025, built mainly for browsers
As of October 2026, computer use is offered by Anthropic, OpenAI and Google, mostly inside agent features and developer platforms, and still advised for supervised, low-risk tasks first.
Questions people ask
What is Claude computer use?
A capability Anthropic introduced in October 2024 that lets Claude operate a computer from screenshots: it chooses where to click, what to type and when to scroll, and the developer's app carries out each action. It is used through Anthropic's developer platform and in Claude's agent products.
Is computer use the same as a browser agent?
A browser agent is computer use limited to a web browser. Most current products focus on browsers, because that is where most tasks happen and it is easier to keep contained. Full desktop control is harder and less mature.
Is it safe to let AI control your computer?
Only with precautions. Run it in a separate virtual machine or browser profile, don't give it passwords or payment details, keep confirmations on, and watch it on unfamiliar sites. Web pages can contain hidden text designed to trick it.
Why doesn't the AI just use an API instead?
When an API exists, that is faster and more reliable, and agents prefer it. Computer use fills the gap for the many websites and older apps that only have a human interface.