Anthropic has written Computer Use into the official document of Claude Code, which is currently a preview function. After opening it, Claude can open the application, click buttons, enter content and view the screen in the CLI. The value is to supplement native GUI testing, and the risk is more immediate because it operates on your real desktop rather than the sandbox.
- What has been updated this time?
- The terminal has a graphical interface operation Computer Use is aimed at native apps, simulators and non-API-free tools. It is suitable for UI verification, recurring layout problems, and running local processes. Claude will first try MCP, Bash, or Chrome extensions, and switch to screen control when the first few are not applicable.
- The ability to be more stringent than normal functions is turned off by default and must be manually enabled through/mcp in interactive sessions. Currently, only macOS is supported, requiring Claude Code 2.1.85 and above, claude.ai account, and Pro or Max subscriptions;Team, Enterprise and-p non-interactive modes are not available.
- How to self-check and see the risks
- Whether self-check has taken effect, first open/mcp in the session to see if computer-use appears and can be enabled; then send a task that requires a GUI to see if the system pops up auxiliary functions and screen recording authorization. Being able to pull up an application and enter approval by application usually means that the entrance is available.
- It is stronger, but the border is also higher. Officials have given several guardrails, including authorization per session per application, only one session is allowed to control the machine globally, pressing Esc to interrupt immediately, and screenshots that do not enter the terminal window. Even so, high-privileged applications still need to be released carefully, because this is not just running commands, but directly touching the real desktop.
