Anthropic computer use
Computer use
Anthropic computer use is a client-side toolset in the Claude API that lets a model see a desktop through screenshots and act with mouse and keyboard calls. The application that sends the request runs every action in an environment it controls, such as a container or virtual machine, and returns the result. Anthropic publishes a reference implementation with a Dockerfile, an agent loop and a web interface in its claude-quickstarts repository.
One tool entry declares the whole toolset; earlier tool versions need beta headers. Anthropic recommends a dedicated virtual machine or container with minimal privileges.

Pinout
- MCP server: not checked yet
- HTTP API: supported, verified
- command line: supported, verified
- Python SDK: supported, verified
- TypeScript SDK: supported, verified
- OpenAPI description: not checked yet
| Interface | Status | Source we read | Read on |
|---|---|---|---|
| MCP server | Not checked yet | — | — |
| HTTP API | Supported, verified | Computer use tool, API usage (tool_use blocks) | |
| Command line | Supported, verified | Computer use tool, quick start (ant messages create) | |
| Python SDK | Supported, verified | Computer use tool, SDKs (Python toolset class, beta) | |
| TypeScript SDK | Supported, verified | Computer use tool, SDKs (TypeScript toolset class, beta) | |
| OpenAPI description | Not checked yet | — | — |
Facts
| Tool type | computer_toolset_20260801; the page says the request needs no beta header. Computer use tool, API usage, read . |
|---|---|
| Member actions | 17 member tools, including screenshot, zoom, clicks, drag, scroll, type and key. Computer use tool, toolset members, read . |
| SDKs and CLI | Python and TypeScript SDKs include a computer toolset class (beta); the quick start also shows an ant messages create command. Computer use tool, SDKs and quick start, read . |
| Reference implementation | anthropics/claude-quickstarts, folder computer-use-demo: Dockerfile, tool implementations, agent loop and web interface. Computer use tool, reference implementation, read . |
| Reference implementation license | MIT. claude-quickstarts repository, license, read . |
| Platforms | Claude API, Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud, Microsoft Foundry. Computer use tool, platform availability, read . |
| Pricing | Billed under standard tool use pricing in USD; declaring the toolset adds about 4,500 input tokens to a request. Pricing page, computer use section, read . |
| Last commit | . GitHub API, latest commit on the default branch of anthropics/claude-quickstarts, read . |
Call example
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."}],
)
print(response)Limits
- The page states the toolset is not currently available in Claude Managed Agents.
- Every action runs in an environment the developer operates; the page recommends a dedicated virtual machine or container.
- Long agent loops accumulate screenshots as input tokens.
Strengths
- A single tool entry declares all member actions.
- A reference container implementation is published under the MIT license.
- The toolset is offered on several cloud platforms besides the Claude API.
Pricing
As written on Anthropic computer use's own page on the read date. Prices change: check there before you pay.
Alternatives to Anthropic computer use
All computer use tools for agents
MCP server that lets AI agents operate Windows applications.

Open-source drivers and sandboxes that let AI agents use computers.

Gemini API tool where the model reads screenshots and returns UI actions.

Simular's open-source framework for computer-use agents, plus a hosted API.
Last check: . Method: public repository and docs read automatically, page and day written down. Next check due within 90 days. How we check
Is this your tool? Write to hello@agent-tools.directory from an address on its domain to correct a pin or a fact.