Anthropic has expanded Claude's computer control capabilities to its Code and Cowork AI agents, allowing the chatbot to autonomously open files, browse the web, and run development tools on macOS devices. The feature requires the Claude desktop app paired with the mobile app, and works by first attempting to use direct API connections to services like Slack and Google Workspace before falling back to controlling your mouse, keyboard, and display directly. Claude asks for explicit permission before taking control, and the functionality is currently limited to Claude Pro and Max subscribers.
This builds on the computer control features Anthropic introduced for Claude 3.5 Sonnet in 2024, but extending it to specialized coding and workplace agents signals their push toward more autonomous AI workflows. However, the "research preview" label and macOS-only limitation reveal how early-stage this technology remains. The fact that Anthropic openly admits "complex tasks sometimes need a second try" and that screen-based control is "slower than using a direct integration" shows they're being refreshingly honest about current limitations.
As I covered when Claude's initial computer control launched, the core challenge remains: giving AI direct system access introduces significant security and reliability risks. Anthropic's approach of prioritizing API connections over screen control is smart, but the fallback to direct mouse/keyboard manipulation still feels like a brute-force solution to integration problems that proper APIs should solve.
For developers, this represents progress toward AI agents that can actually execute multi-step workflows autonomously. But the research preview status and acknowledged reliability issues mean this is still experimental territory. If you're building production systems, stick with structured APIs and webhooks rather than betting on AI that clicks around your screen.