Computer Use

0:54 Computer use lets Claude control a computer the way a person does: it looks at a screenshot, decides where to click or what to type, and asks your program to do it, one step at a time. On the Claude API it is a single tool entry, {"type": "computer_toolset_20260801"}, which gives Claude 17 member tools such as screenshot, left_click, type and zoom (as of Nov 6, 2026). Anthropic never touches your machine: your application takes the screenshots and performs the clicks, so computer use is a client tool in the sense of Tool Use.

0:00 The feature first appeared on Oct 22, 2024 as a public beta with the upgraded Claude 3.5 Sonnet, on the API, Amazon Bedrock and Google Cloud's Vertex AI. Anthropic called it "at times cumbersome and error-prone" at the time. It left beta on Aug 19, 2026, when the toolset version replaced the older single-tool versions and a separate browser use tool arrived beside it.

Quick facts

How does computer use work?

1:26 Computer use is an agent loop that your code runs:

1. Send a request with the toolset in tools and a task in the user message.
2. Claude replies with one or more member tool calls (for example a screenshot, then a left_click at given coordinates) and stop_reason: "tool_use".
3. Your application performs each action in order on a real or virtual screen and returns one tool_result per call: an image for screenshot and zoom, a short text such as "OK" for the rest.
4. Claude looks at the new screen and asks for the next actions, until it answers in text with no tool calls.

The docs' own example declares the toolset next to the bash and text editor tools, which Claude often uses alongside it:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    tools=[
        {"type": "computer_toolset_20260801"},
        {"type": "text_editor_20250728", "name": "str_replace_based_edit_tool"},
        {"type": "bash_20250124", "name": "bash"},
    ],
    messages=[{"role": "user", "content": "Save a picture of a cat to my desktop."}],
)

No beta header is needed for the toolset. Claude can return several actions at once (click, type, then screenshot); the docs ask you to run them in order, stop at the first failure and mark every later action as not executed (Computer use tool).

The 17 member tools

The toolset's members are screenshot, zoom, left_click, right_click, middle_click, double_click, triple_click, left_click_drag, mouse_move, left_mouse_down, left_mouse_up, cursor_position, scroll, type, key, hold_key and wait (Computer use tool). zoom takes a rectangle and returns that part of the screen at full resolution, which helps Claude read small text such as file names or status bars; if your environment cannot produce it, turn it off with "configs": {"zoom": {"enabled": false}}, which saves about 410 input tokens.

Which models support it?

Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5 and Claude Fable 5.1 use the toolset, and Claude 5.5 models return a 400 error for computer_20251124 on the Claude API (Release notes). Opus 4.7, Opus 4.6, Sonnet 4.6 and Opus 4.5 support computer use only through computer_20251124 with its beta header, and Amazon Bedrock keeps that version (Computer use tool).

Screen size and coordinates

The docs suggest 1024x768 or 1280x720 for general desktop tasks and 1280x800 or 1366x768 for web apps, and warn against going above 1920x1080. Claude 4.7 and later models accept screenshots up to 2,576 pixels on the long edge. The API does not shrink oversized images for you: resize screenshots yourself and scale Claude's coordinates back to the real screen, and on a Mac Retina display halve them (Computer use tool).

Anthropic's reference implementation, with a container, the tool code, an agent loop and a web interface, is the computer-use-demo quickstart on GitHub.

Safety

Computer use carries risks that ordinary chat does not, and Anthropic's docs say so directly: Claude may follow instructions it finds on a web page or in an image, even when they conflict with yours (prompt injection). The docs recommend four precautions (Computer use tool):

1. Run it in a dedicated virtual machine or container with minimal privileges.
2. Keep sensitive data, such as account logins, out of its reach.
3. Limit internet access to an allowlist of domains.
4. Have a person confirm anything with real consequences, such as payments, accepting terms or cookie banners.

Anthropic also runs classifiers over screenshots that look for prompt injection and steer Claude to check whether an instruction really came from you; customers can opt out through support. Products that offer computer use are asked to tell end users about the risks and get their consent first.

Benchmarks

The usual yardstick is OSWorld, a set of real desktop tasks. At launch, Anthropic reported 14.9% on OSWorld's screenshot-only category for Claude 3.5 Sonnet, against 7.8% for the next best system. Anthropic's newer posts use later OSWorld releases that are not comparable with those early numbers: it reports 81.8% on OSWorld 2.1 (partial) for Opus 5.5 and 80.1% for Sonnet 5.5, and gave no competitor scores on that version. These are Anthropic's own measurements; see Claude Benchmarks for independent leaderboards.

Computer use or browser use?

For work that happens only in a web browser, the browser use tool (browser_toolset_20260801, launched Aug 19, 2026) is the narrower option: your application hosts the browser, and Claude reads the page's accessibility tree, elements, forms and tabs instead of guessing from pixels. It costs about 6,600 input tokens with default members. Computer use remains the choice for desktop apps and anything outside a browser. Inside the Claude apps, Claude in Chrome and Cowork offer these abilities without code.

Frequently asked questions

Is Claude computer use free? Claude computer use has no separate fee on the API: you pay normal token prices, plus about 4,500 input tokens for the toolset and image tokens for every screenshot (Pricing). See Claude API Pricing.

Does Claude computer use control my own computer? Only if your application lets it. Through the API, Claude only asks for actions, and your program decides whether and where to perform them; Anthropic recommends a separate virtual machine or container.

Is Claude computer use still in beta? No. On the Claude API, computer use has been generally available as computer_toolset_20260801 since Aug 19, 2026; in the Claude desktop apps (Cowork and Claude Code), computer use is still a research preview as of Nov 6, 2026.

What is the difference between computer use and the browser use tool in Claude? Claude computer use works from screenshots of a whole desktop, while the browser use tool drives a browser through its page structure, which is usually faster and more reliable for web-only tasks.

Can Claude computer use run on Amazon Bedrock? Yes. Amazon Bedrock supports Claude computer use, but it keeps the older computer_20251124 version rather than the toolset (Computer use tool).

Sources

• Claude | Computer use for automating operations, Anthropic, Oct 22, 2024 (video)
• Claude | Computer use for coding, Anthropic, Oct 22, 2024 (video)
• Computer use tool, Anthropic docs, read Nov 6, 2026
• Tool reference, Anthropic docs, read Nov 6, 2026
• Pricing, Anthropic docs, read Nov 6, 2026
• Release notes, Anthropic docs, read Nov 6, 2026
• Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku, Anthropic, Oct 22, 2024
• Introducing Claude Opus 5.5, Anthropic, Sep 22, 2026
• Introducing Claude Sonnet 5.5, Anthropic, Sep 28, 2026
• Release notes, Claude Help Center, read Nov 6, 2026
• computer-use-demo, Anthropic on GitHub, read Nov 6, 2026