Claude Computer Use: What It Is and How to Use It (2026)

Last Updated: November 6, 2026

Summary: Claude computer use is Anthropic's feature that lets Claude operate a computer the way a person does, by reading screenshots and then moving the mouse, clicking and typing. Developers get it through the Claude API's computer_toolset_20260801 toolset, which needs no beta header, and Pro, Max and Ultra subscribers can turn it on in Cowork and Claude Code in the Claude Desktop app on macOS and Windows. API use is billed as normal tokens, and Anthropic advises running it in an isolated environment with a human approving anything consequential.

Most AI tools can only work with software that offers an API. Claude computer use removes that limit: if a person can do a task by looking at a screen and clicking, Claude can attempt it. That makes it useful for legacy desktop software, internal tools with no integrations and testing apps you are building.

This guide covers availability, setup, cost, limits, safety and how it compares with OpenAI, checked against official documentation in November 2026.

Claude computer use is a capability that lets Claude control a desktop through screenshots plus mouse and keyboard actions. Anthropic's computer use documentation describes a tool that "provides screenshot capabilities and mouse/keyboard control for autonomous desktop interaction."

Claude computer use does not plug into an app's code. Claude sees the screen as an image and requests an action, such as a click at specific pixel coordinates or a line of typed text. Software on your side performs the action, takes a fresh screenshot and sends it back, and the cycle repeats until the task is done.

Anthropic first released computer use as a public beta on October 22, 2024, with the upgraded Claude 3.5 Sonnet, according to PPC Land's report on the launch. On August 19, 2026, the API version came out of beta as the computer_toolset_20260801 toolset, per Anthropic's platform release notes. It is a clear example of an AI agent, software that takes actions toward a goal; our explainer on what AI agents are covers the wider category.

Claude computer use is available in the Claude API, in Cowork and Claude Code inside the Claude Desktop app, and in the Claude Code command-line tool on macOS. Claude in Chrome is a related but separate feature that controls only your browser.

The consumer version arrived in March 2026 as a research preview for Pro and Max subscribers on macOS, according to Engadget. Anthropic's help center article on computer use in Cowork now lists it as a beta on macOS and Windows for Pro, Max and Ultra, and says Team and Enterprise plans "don't have access to computer use at this time."

Claude in Chrome is a browser extension. Anthropic's Claude Code Chrome documentation says it opens new tabs and shares your browser's login state, so Claude can work in any site you are already signed into.

Claude computer use works as a loop: Claude asks for an action, your environment performs it and returns the result, and Claude decides the next step. Anthropic's docs describe four steps:

Anthropic calls steps 3 and 4, repeated without user input, the agent loop. The current toolset supports batch actions, so Claude can request several actions in one turn. It has 17 actions in total, including five kinds of click, dragging, scrolling, typing, key presses and zoom, which returns a chosen screen region at full resolution so Claude can read small text.

To use Claude computer use on your own computer, enable one toggle in the Claude Desktop app. Anthropic's help center lists these steps:

In Cowork, Claude tries connectors first (such as Gmail, Google Drive, Microsoft 365 or Slack), then the browser, and only then direct screen control, which Anthropic calls "slower and more error-prone." Investment, trading and cryptocurrency apps are blocked by default, and you can add your own apps to a blocklist. Your computer must stay awake with the Desktop app open.

In the Claude Code terminal tool, computer use is a built-in MCP server called computer-use that is off by default. Anthropic's Claude Code computer use guide calls it a research preview on macOS for Pro and Max plans.

Claude Code prefers an MCP server, then Bash, then Claude in Chrome, and uses screen control only for native apps, simulators and tools with no API. New to the tool? Start with our guide to what Claude Code is.

On the Claude API, add {"type": "computer_toolset_20260801"} to the tools array of a Messages API request. Anthropic's documentation states that "the request needs no beta header" for this toolset. Its quick start pairs it with the text editor tool (text_editor_20250728) and the bash tool (bash_20250124). Zoom is on by default; turn it off with "configs": {"zoom": {"enabled": false}}.

The older computer_20251124 tool takes display settings such as display_width_px and display_height_px. Anthropic's release notes say newer models on the Claude API and Google Cloud return a 400 error for it, while it keeps working on Bedrock. The toolset is not supported on several older models, including Claude Opus 4.7, Opus 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5.

You still build the environment Claude controls. Anthropic's computer-use-demo reference implementation runs a virtual desktop in Docker with an agent loop and web interface. The Python SDK (1.12.0) and TypeScript SDK (0.132.0) also added computer toolset classes, in beta, that run the loop for you.

Claude computer use has no separate fee. Anthropic's pricing page says it "follows the standard tool use pricing," so you pay normal token rates for your model, plus these extras:

A worked example. This is an illustration built from Anthropic's figures, not a measured result. Each turn of the agent loop resends the conversation so far. If a task takes 20 turns on Claude Sonnet 5.5 and each turn adds one 1,500-token screenshot, requests grow from about 6,000 to about 34,500 input tokens. That totals roughly 405,000 input tokens, or about $0.81, before output tokens. Prompt caching, billed at 0.1x the input rate for cache hits, can cut that sharply. Our Claude API pricing guide explains caching and batch discounts.

Claude computer use is strongest on tasks no API can reach and weakest on tasks needing speed or perfect precision. Anthropic recommends it for "background information gathering, automated software testing" in trusted environments, and Claude Code's docs add testing native apps, reproducing layout bugs and driving GUI-only tools such as the iOS Simulator.

Anthropic's documentation lists these limits:

The practical rule: if a connector or API exists, use that instead.

Claude computer use carries real risk because Claude acts on live software and sees anything on screen. Anthropic's computer use safety guide states that in the Desktop app there is no sandbox between Claude and your apps.

The biggest threat is prompt injection: Anthropic warns that "instructions on webpages or contained in images might override your instructions." The API runs classifiers on screenshots to flag this, but Anthropic says its other precautions still matter. For developers, it advises:

For Desktop users, Anthropic advises not granting access to banking, healthcare or government apps, starting with simple tasks and closing sensitive files first. The API tool is listed as eligible for Zero Data Retention and HIPAA, and screenshots stay in your environment rather than with Anthropic.

OpenAI offers computer use as a computer tool in its Responses API for developers and as ChatGPT agent for consumers. The main difference is location: Claude's consumer feature works on your own Mac or PC, while ChatGPT agent works on a virtual computer OpenAI runs.

According to OpenAI's API changelog, built-in computer use arrived in GPT-5.4 through the Responses API on March 5, 2026, then came to GPT-5.4 mini and GPT-5.5. On September 29, 2026, OpenAI added computer use to its Agents API, where agents work in an OpenAI-hosted browser. OpenAI's computer use guide recommends network rules, an isolated browser profile or desktop, and approval before changes.

ChatGPT agent launched on July 18, 2025, combining Operator's browser actions with deep research on its own computer, as Windows Central reported; at launch it was available on ChatGPT Pro, Plus and Team. For a wider comparison of the two assistants, see our ChatGPT vs Claude guide.

The right option depends on whose computer Claude should control.

What is Claude computer use?
Claude computer use is an Anthropic feature that lets Claude operate a computer by viewing screenshots and controlling the mouse and keyboard. Developers use it through the Claude API, and subscribers can enable it in Cowork and Claude Code in the Claude Desktop app.

How much does Claude computer use cost?
There is no separate fee. On the API it is billed as standard token usage, with about 4,500 extra input tokens per request for the toolset and roughly 1,000 to 1,800 tokens per screenshot. In the Desktop app it is part of Pro, Max and Ultra plans.

Does Claude computer use need a beta header?
No, not for the current computer_toolset_20260801 toolset on the Claude API. The older computer_20251124 tool, still used on Amazon Bedrock, needs the computer-use-2025-11-24 beta header.

Does Claude computer use work on Windows?
Yes, in the Claude Desktop app, where Cowork and Claude Code support computer use on macOS and Windows. The Claude Code CLI version is macOS only.

What is the difference between Claude computer use and Claude in Chrome?
Claude computer use controls desktop apps through screenshots, clicks and typing. Claude in Chrome is a browser extension that works only inside Chrome or other Chromium browsers and uses your existing logins, which suits web tasks.

Claude computer use lets Claude work with any software a person can see and click, filling the gap that APIs and connectors leave. Developers get a toolset that is out of beta and needs no header, while Pro, Max and Ultra subscribers can switch it on in the Claude Desktop app.

Treat it as the tool of last resort. It is slower and less precise than a connector and acts on real software, so isolate it where you can, approve consequential actions yourself and start small. For more on Claude's growth, see our Claude AI statistics.

Written by the AI Business Weekly Editorial Team. Reviewed, edited and published by Sameer Khan. This article was AI-assisted, then reviewed by Sameer Khan before publishing.