Anthropic Launches ‘Computer Use’ Feature for Claude AI, Enabling Automated Computer Control for Developers
Anthropic recently unveiled a new ‘computer use’ feature for its Claude AI model. This capability allows Claude to directly control a computer’s interface by simulating human mouse and keyboard actions, enabling more advanced automation tasks. Developers can now access this beta feature via API, which is primarily optimized for software engineering scenarios.
Core Feature Mechanics
Claude’s ‘computer use’ feature is built on visual understanding and action planning. The model first analyzes the current interface state through screenshots and then generates a sequence of actions like mouse clicks, drags, and keyboard inputs based on natural language instructions. For example, Claude can browse web pages, fill out forms, or navigate desktop applications without relying on specific APIs or structured data.
Anthropic stated in its official blog that this feature builds on existing research in computer use and has been extensively optimized with reinforcement learning. It currently supports the Claude 3.5 Sonnet model, with significant improvements in operational precision and speed. Developers only need to enable the computer_use tool in their API calls to activate this capability.
Optimized for Software Engineering
This feature is particularly suited for automating development tools, such as in code review, bug fixing, and testing workflows. Claude can directly navigate within an IDE, edit files, or run command-line instructions, helping developers handle repetitive tasks. Anthropic’s tests show that Claude’s performance on software engineering benchmark tasks surpasses models that previously relied solely on text-based interfaces.
According to Anthropic, the feature is already being used internally to build AI agent prototypes and is planned to be extended to more model versions. Developer feedback indicates that it significantly boosts productivity, especially in legacy systems or environments without API support.
Security and Access Control
To ensure safety, Anthropic has built in multiple protective mechanisms. Before each operation, users must manually confirm critical actions; the model runs in a sandboxed environment, restricting access to system files; and it also supports rate limiting and anomaly detection. Anthropic emphasizes that access is currently limited to API users and is not supported on the Claude.ai web interface.
Several tech media outlets, such as TechCrunch and The Verge, have already reported on this update. Anthropic’s official documentation provides detailed integration guides, allowing developers to start testing immediately. In the future, this feature is expected to drive the application of AI in real-world workflows, marking a step forward for large models toward becoming general-purpose agents.
