Claude Computer Use
anthropic.com
- Linux
- API
What Claude Computer Use is
Claude Computer Use lets Claude operate a computer interface by reading screenshots and requesting actions such as clicks and typing. Developers add the computer-use tools to a Messages API request, run those actions in an environment they control, then return screenshots or other results so Claude can continue until a task is complete. The toolset includes screenshot, click, typing, and other actions, with short sequences supported. Anthropic describes uses including repetitive-process automation, software building and testing, and open-ended research. Availability is listed through the Claude API and Google Cloud, with Amazon Bedrock announced at launch; it is not currently available in Claude Managed Agents. The reference setup uses a Linux desktop, virtual X11 display, Firefox, and LibreOffice, isolated in Docker. Pricing follows standard tool-use pricing: model tokens are charged, and screenshots count as image input. Anthropic recommends low-risk tasks, human confirmation for consequential actions, and minimal-privilege containers or virtual machines. Scrolling, dragging, zooming, coordinate selection, speed, and niche workflows can be unreliable.
Who it suits
Claude Computer Use suits developers building applications that need API-driven desktop or browser control. It is intended for controlled environments where human oversight and isolation can be provided.
Good to know, and what to watch
- Screenshot-based computer interaction
- Supports clicks, typing, and short action sequences
- API and Google Cloud availability
- Docker reference environment for isolation
- Human confirmation guidance for consequential actions
- No free plan
- Model tokens and screenshot image input incur costs
- Scrolling, dragging, and zooming may be unreliable
- Not available in Claude Managed Agents
The verdict
Claude Computer Use provides an API workflow for letting Claude act inside a controlled computer environment. Start with low-risk tasks and add isolation and approval controls before handling consequential actions.
What the maker publishes
- Accuracy limits
- Anthropic advises against using computer use for tasks requiring perfect precision or sensitive user information without human oversight. source
- Actions
- The tool can request actions including screenshots, clicks, and typing, and can return short sequences of actions together. source
- Availability
- The current computer toolset is available on the Claude API and Google Cloud; the page says it is not currently available in Claude Managed Agents. source
- Capabilities
- The current computer toolset has 17 member tools, including screenshot, left_click, type, and zoom. source
- Company
- Anthropic describes itself as an AI safety and research company that builds reliable, interpretable, and steerable AI systems. source
- Data handling
- Screenshots, mouse actions, keyboard inputs, and session files are stored in the developer's environment; Anthropic processes screenshots and action requests in real time as part of API calls. source
- Data retention
- Anthropic says the client-side computer use tool is eligible for zero data retention, with API request retention governed by its API data retention terms. source
- Developer implementation
- Developers add the computer use toolset to a Messages API request and run Claude’s requested actions in their own environment. source
- Environment
- Anthropic’s reference environment uses a Linux desktop with a virtual X11 display and applications including Firefox and LibreOffice. source
- Headquarters
- Anthropic’s transparency hub identifies San Francisco, California, as its headquarters. source
- How it runs
- Your application runs each computer action in an environment you control, returns the result to Claude, and repeats the process until the task is complete. source
- Integrations
- At launch, Anthropic said computer use was available through its API, Amazon Bedrock, and Google Cloud Vertex AI. source
- Isolation
- The reference implementation runs the computer-use environment inside a Docker container for security and isolation. source
- Limitations
- Anthropic describes computer use as imperfect and says scrolling, dragging, and zooming can be challenging. source
- Maker identity
- Anthropic describes itself as an AI safety and research company that builds reliable, interpretable, and steerable AI systems. source
- Maker security certifications
- Anthropic's Trust Center lists SOC 2 Type II, ISO 27001:2022, ISO/IEC 42001:2023, and HIPAA among its compliance resources. source
- Performance limits
- Anthropic cautions that computer use can be slow, coordinate selection and tool choice can be unreliable, and performance may be lower in niche or multi-application workflows. source
- Pricing
- Computer use follows standard tool-use pricing; requests incur model token costs, and screenshots are billed as image input. source
- Prompt injection protection
- Classifiers scan tool results such as screenshots for potential prompt injections and can steer Claude to verify whether an instruction came from the user. source
- Prompt injection safeguards
- Anthropic says classifiers scan tool returns for potential prompt injections and steer Claude to check whether suspicious instructions came from the user. source
- Safety advice
- Anthropic recommends starting with low-risk tasks because computer use can be error-prone and may create risks such as spam, misinformation, or fraud. source
- Security guidance
- Anthropic recommends using a dedicated virtual machine or container with minimal privileges, restricting internet access, and requiring human confirmation for consequential actions. source
- Task loop
- The application returns screenshots and other tool results to Claude so it can continue acting until it completes the task. source
- Use cases
- Anthropic describes uses including automating repetitive processes, building and testing software, and conducting open-ended research. source
Checked against the maker's pages on 4 Oct 2026.