Anthropic's Claude: Code & Cowork Now Controls Your Computer

Anthropic has announced a significant enhancement to Claude, enabling its Code and Cowork AI tools to autonomously perform tasks directly on your computer. This innovative feature empowers Claude to automatically open files, navigate web browsers, utilize various applications, and execute developer tools “with no setup required,” even when users are away from their devices.

Currently available as a research preview, these advanced capabilities are exclusively for Claude Pro and Max subscribers. Anthropic notes that computer usage for this feature is presently limited to macOS devices, though this may change. The update builds upon the autonomous functionalities first introduced with Claude’s 3.5 Sonnet AI model in 2024, now extending these sophisticated interactions to the chatbot’s specialized Code and Cowork AI agents, particularly benefiting programmers.

To activate this functionality, users must have the Claude desktop application running on a supported macOS device, paired with the chatbot’s mobile app. The system prioritizes establishing connections with supported services like Slack and Google Workspace applications. However, should a direct connector not be available, Claude is designed to execute tasks by directly controlling the user's browser, mouse, keyboard, and display. Crucially, Anthropic emphasizes that Claude will “always ask for your explicit permission” before initiating actions such as exploring, scrolling, or clicking to complete a task.

Anthropic highlights that this feature integrates “especially well” with Dispatch, a recently launched cross-device capability allowing users to assign tasks to the Claude desktop app directly from their mobile phones. Despite its potential, the company transparently acknowledges that the feature is not yet flawless. In its announcement blog, Anthropic stated, “Complex tasks sometimes need a second try, and working through your screen is slower than using a direct integration.” The company is releasing this preview early to gather user feedback, aiming “to learn where it works and where it falls short.”