Toolofficial publisher

computer use (Anthropic)

Tool : A precise action the AI can call · requires anthropic-api

Lets Claude see the screen, move the mouse, and type on the keyboard on a machine.

The most dangerous capability in the catalogue: never without an isolated VM and supervision.

Docs

Overall score

5.3 /10

Security: serious reservations

Utility

7.0

Reliability

6.0

Security

Serious reservations

Ease

Technical · Technical

Security reflects where data goes, requested permissions and incidents on record at the verification date. It is not a guarantee of zero risk: check the deployment mode and the connections you enable.

What it does, concretely

Some computing tasks are too complex to describe in words: navigating a graphical interface, filling out a form, moving files on a desktop. Computer use lets Claude act directly on a desktop environment by seeing the screen and controlling the mouse and keyboard, much like a human user would.

In practice, the tool exposes 17 member actions to Claude (screenshot, left click, keyboard input, zoom, etc.) through a single type identifier in the API. Claude receives screenshots, decides which actions to chain together, sends them as blocks, and your application executes them in the environment you control before returning the results. This loop repeats until the task is complete.

This tool is aimed at developers building agents capable of automating desktop tasks: software testing, data extraction from graphical interfaces, automation of repetitive workflows on Linux. It requires familiarity with the Anthropic API and the setup of an isolated execution environment.

It is not suited for non-technical users, nor for use cases limited to web pages: for those, the browser use tool is a better fit and does not require a full desktop environment. It is also not available in Claude Managed Agents.

What it offers

·Automatic screenshot: Claude takes screenshots to observe the state of the desktop before each decision. (paid)
·Mouse and keyboard control: click, double-click, text input, scrolling, and zoom are available among the 17 member actions. (paid)
·Batch actions: Claude can plan and send multiple actions in a single response, executed in order. (paid)
·Prompt injection detection: classifiers automatically analyze screenshots to flag injection attempts and request human confirmation before proceeding. (paid)
·Compatibility with bash and text editor tools: computer use is commonly combined with these tools for tasks mixing desktop and command-line work. (paid)
·Available on the Claude API and Google Cloud via the computer_toolset_20260801 identifier. (paid)
·Backward compatibility: integrations based on the older computer_20251124 version continue to work. (paid)

Our tips

Always add the bash and text editor tools alongside computer use: Claude naturally uses them as complements for complex desktop tasks.
If your application loop does not end with a screenshot, manually attach one as an image block on the last result in the batch: this avoids an extra round trip with the API.
Process batch actions in strict order and stop at the first failure; still return a tool_result block for each unexecuted action with the exact prescribed text, otherwise the API rejects the request.
If your use case does not include a human in the loop, you can disable the injection classifier protection by contacting Anthropic support, but isolation precautions remain essential.
For tasks limited to web pages, prefer the browser use tool: it acts directly on page content without requiring a full desktop environment.

Watch out

Claude may follow instructions found in on-screen content (web pages, images) even if they contradict your initial instructions: environment isolation and human supervision remain necessary even with the classifier protection enabled.
Batch actions execute sequentially, not in parallel: an action that depends on a previously failed one will not run, which can interrupt complex workflows if error handling is not properly implemented.

In practice

Price: Paid
Commercial use: Commercial use: yes
Available on: api
Host: anthropic-api
FR interface: No

Security and data

Rating: Security: serious reservations
Capabilities: readswritesexecutesnetwork
Your data: Local or cloud depending on use
Maintenance: Active

Links and repository checked on 25 October 2026 · added on 1 October 2024

How to install computer use (Anthropic)

1. Obtain an Anthropic API key at platform.claude.com. 2. Set up an isolated virtual machine or container with minimal privileges to execute desktop actions. 3. Install the Anthropic SDK for your language (e.g. pip install anthropic for Python). 4. Add {"type": "computer_toolset_20260801"} to the tools array of your Messages API request, with no additional beta header. 5. Implement the agent loop: execute each tool_use block in order, return one tool_result per block, and repeat until stop_reason is no longer tool_use. 6. Never expose sensitive data (credentials, confidential files) from within the execution environment.

pip install anthropic

Official documentation ↗

In the same category

Whole category Automation & agents →