Anthropic Unveils Claude's Groundbreaking Computer Control Capability
In a significant leap forward for artificial intelligence, Anthropic has officially launched a new feature that empowers its Claude AI to directly operate users' computers. This expansion, arguably the most consequential yet, enables Claude to move cursors, click through menus, open applications, fill out forms, and navigate browsers in a manner indistinguishable from human interaction. The feature, dubbed "computer use," is now live as a research preview exclusively for Claude Pro and Max subscribers on macOS platforms.
How Claude's Computer Use Feature Functions
The operational premise of this innovation is both straightforward and transformative. Users simply describe a desired outcome, and Claude systematically works backward through every necessary step to achieve it on the actual machine, all while the user can attend to other tasks. Claude does not immediately seize control of the mouse and keyboard; instead, it first checks for existing connectors to services like Slack and Google Calendar. If a direct integration is available, Claude utilizes that pathway. Only in the absence of such connectors does it resort to controlling the screen directly.
From that point, Claude can perform a wide array of actions, including opening files, browsing the web, typing into forms, and navigating menus. Crucially, the AI always seeks permission before accessing a new application, and users retain the ability to halt operations at any moment. This feature synergizes with Dispatch, Anthropic's task-assignment tool, which allows users to delegate work from their mobile devices. The combination enables scenarios where tasks assigned during a commute are completed and ready on the desktop upon arrival, with support for recurring assignments like daily inbox scans or weekly metric pulls.
Security Considerations and Early-Stage Limitations
Anthropic has been transparent about the inherent risks associated with granting an AI model access to a computer. The primary concern is prompt injection, where hidden instructions in webpages or documents could surreptitiously override user commands. To mitigate this, Anthropic has implemented a real-time classifier that scans for such threats, prompting Claude to pause and seek user confirmation before proceeding. Additionally, certain sensitive applications are disabled by default as a precautionary measure.
The company acknowledges that this is early-stage software, with limitations including occasional need for retries on complex tasks, slower performance when navigating screens compared to direct integrations, and reduced reliability with niche or unfamiliar software. These factors underscore the measured approach of the research preview, aimed at identifying shortcomings before a broader rollout.
Competitive Landscape and Anthropic's Strategic Position
Claude enters the increasingly crowded AI agent space relatively late, facing direct competition from frameworks like OpenClaw, an open-source agentic system that has fostered a vibrant ecosystem of semi-autonomous AI tools. Recent entries, such as Nvidia's NemoClaw with built-in security guardrails, further intensify the competition. Anthropic distinguishes itself with a focus on tighter safety controls, a polished consumer product rather than a developer toolkit, and deeper integration with its own app ecosystem. This represents a more cautious bet in a fast-moving field, raising questions about whether this trade-off will resonate with users.
Getting Started with Claude's Computer Use
To access this feature, users must be Claude Pro or Max subscribers with the Claude desktop app installed and running on macOS. The mobile app alone does not support computer control but can be used with Dispatch to assign tasks that execute in the background on the desktop. Anthropic frames this release as a research preview to gather insights and refine the technology, a prudent approach given the direct access to live desktops.
