Claude launches “Computer Operation” function: can directly control MacOS applications

[ Gearbest Technology News]On March 31, foreign media reported that Anthropic announced that the “computer use” function of the Claude model was officially integrated into the Claude Code CLI, and a research preview was opened to MacOS users. Users can enable the built-in MCP server in the command line interface, allowing Claude to open applications independently, simulate clicking and typing, view the screen in real time, and truly “hands-on” graphical interface tasks. For example, Claude can write Swift code, compile and launch the MacOS menu bar application, and click on the controls one by one to verify, without manual intervention.

This function follows a clear task priority: first use the corresponding MCP server, secondly call Bash or Chrome extension, and finally enable the “Computer Use” function. Its application scenarios include end-to-end UI testing, debugging visual layout issues, and operating proprietary software without command lines or APIs (such as design tools, hardware control panels, etc.), which greatly expands the automation capabilities of software development and testing.

In terms of security, Anthropic adopts careful design. When Claude operates an application for the first time, a prompt will pop up on the terminal to clarify the application name, requested permissions and scope of influence. The user can allow the session or deny it. Different application categories have differentiated control levels: browsers and trading platforms are read-only, terminals and IDEs are click-only, and other applications are fully controlled. The system also has built-in mechanisms such as sentinel warning and Esc key emergency stop to ensure that users can intervene at any time.

Currently, this function only supports MacOS and requires Claude Code v2.1.85 or above. It is for Pro and Max package users to use in interactive sessions. Non-interactive mode is not supported yet. Anthropic stated that this feature is still in the research preview stage and will be gradually improved and expanded to more platforms based on feedback. This progress marks the move of AI from “thinking and generating” to “perceiving and operating”, bringing new possibilities for automated testing, application development and GUI tool integration.