Computer use is Anthropic's API capability that gives Claude screenshot vision plus mouse and keyboard control to operate desktop applications autonomously.
Anthropic Computer Use is a beta tool in the Claude API that gives Claude the same interface a person uses: a screenshot of the screen, plus the ability to move the mouse, click, and type. Instead of calling a purpose-built API for each app, Claude looks at a screenshot, decides where to click or what to type next, and your application executes that action inside a sandboxed virtual machine before returning the result. Because it works at the level of pixels and keystrokes rather than app-specific integrations, computer use lets Claude operate almost any desktop software, from spreadsheets to internal tools with no API at all, which matters for teams automating workflows that were never built to be automated.
Computer use runs as an agent loop between your application and the Claude API. Claude never touches your machine directly.
Screenshot capture: your application takes a screenshot of the sandboxed display and sends it to Claude as an image.
Action request: Claude reads the screenshot and responds with a structured tool call, such as clicking coordinates or typing a string of text.
Execution: your code runs that action inside the virtual environment, a Linux desktop with a virtual X11 display, a window manager, and pre-installed apps like a browser and file manager.
Result and repeat: your application takes a new screenshot, sends it back as the result of the tool call, and Claude decides on the next action or ends the loop once the task is done.
Safety classifiers: Anthropic runs automatic checks on incoming screenshots and pauses the loop to ask for human confirmation if it detects a likely prompt injection, hidden instructions embedded in a webpage or image that try to hijack the agent.
Because Claude reasons from pixels rather than a fixed script, it can adapt to interfaces it has never seen. That same flexibility means it can also misread a screen or click the wrong element, so most production setups run it inside an isolated VM with no access to sensitive accounts.
Legacy software automation: an agent that operates a desktop accounting tool or internal admin panel that only ships a GUI, with no API to call directly.
QA testing across full applications: a test agent that exercises a whole desktop app, not just a web frontend, catching regressions a browser-only tool would miss.
Cross-application workflows: an agent that pulls data out of a PDF viewer, pastes it into a spreadsheet, then sends the result from a desktop mail client, chaining apps the way a person would.
Onboarding and training environments: a virtual assistant that walks through a piece of enterprise software step by step inside a sandbox, useful for generating documentation or training material automatically.
Research and background automation: long-running tasks like data entry or repetitive form-filling where speed is not critical, run unattended in a virtual machine and checked afterward.
No, computer use is billed like any other Claude API tool call, at your model's standard per-token rate, plus a small system prompt overhead of around 466 to 499 tokens and roughly 735 tokens per tool definition. There is no separate computer use fee. The bulk of the cost usually comes from the screenshot images Claude processes at each step, since every screenshot consumes vision tokens.
It carries real risk, since Claude reads the screen visually rather than following a fixed script, so hidden instructions on a webpage or in an image can attempt to redirect its actions, a problem known as prompt injection. Anthropic runs classifiers that pause for human confirmation when they detect a likely injection, and recommends running computer use in an isolated virtual machine with no access to sensitive accounts or data.
Browser Use and similar tools only control a browser tab, reading the DOM and clicking web elements. Claude's computer use tool controls an entire desktop screen through pixel-based vision, so it can operate any application, a spreadsheet, a terminal, a native app, not just websites. That broader scope makes it more flexible but generally slower and less accurate on web-only tasks than a browser-specialized tool.
Stagehand is an open-source browser automation SDK that replaces CSS selectors with natural language instructions, built by Browserbase on top of Playwright.
Browser AgentBrowser Use is an open-source Python library that lets AI agents control a real web browser to click, type, and extract data using natural language.
Browser AgentUpdates from the AI world — what shipped, what we’re using in production, and what’s worth your attention. Two emails a month, no spam.
