Explained: Anthropic's new Claude tool that can control your computer; Here's how it works, what it can do, and more

Anthropic has rolled out what is arguably the most consequential expansion of Claude's capabilities yet—the ability to directly operate your computer. Not through a plugin, not through an API call, but by actually moving your cursor, clicking through menus, opening apps, filling forms, and navigating your browser the way a person would. The feature, called computer use, is now live as a research preview for Claude Pro and Max subscribers on macOS, and it arrives as the AI agent space gets increasingly crowded and competitive.The pitch is straightforward but significant: you describe an outcome, and Claude works backwards through every step to get there—on your actual machine, while you're doing something else entirely.Claude uses your apps, and your screen when it can'tClaude doesn't reach for your mouse and keyboard right away. It first checks whether a connector exists for the job—Slack, Google Calendar, and a handful of other services are already wired in. If a connector handles it, that's the path it takes. Only when there's no direct integration does it fall back to controlling your screen.From there, it can open files, navigate a browser, type into forms, and click through menus—anything you'd do yourself. It always asks permission before touching a new application, and you can pull the plug at any point.The feature pairs with Dispatch, Anthropic's task-assignment tool that lets you delegate work from your phone. The combination is what makes the whole thing more than a novelty. Assign a task on your commute, and the finished output is waiting on your desktop when you sit down. You can also set recurring tasks—scan my inbox every morning, pull the weekly metrics every Friday—and Claude handles them on schedule without being asked again.Anthropic's own security warnings are worth readingHanding an AI model the keys to your computer introduces risks that Anthropic doesn't try to downplay.The main one is prompt injection—where instructions buried in a webpage or a document could quietly override what you asked Claude to do. Anthropic has built a classifier that scans for this in real time and asks Claude to pause and confirm with you before proceeding. Certain sensitive apps are also switched off by default.Still, the company is candid that this is early-stage software. Complex tasks sometimes need a second attempt. Claude is slower when navigating through your screen than when using a direct integration—there's no getting around that. And its reliability drops when it encounters niche or unfamiliar software.Claude is entering the AI agent arms race late, but with a different betComputer use lands Claude in direct competition with OpenClaw, the open-source agentic framework that took off earlier this year and built an entire ecosystem of AI tools that can act semi-autonomously on your computer. Nvidia jumped in last week with NemoClaw, a framework for deploying OpenClaw agents with security guardrails baked in.Anthropic's approach is different—tighter safety controls, a polished consumer product rather than a developer toolkit, and deeper ties to its own app ecosystem. It's a more cautious bet in a space where developers are moving fast and loose. Whether that tradeoff wins users over is the real question.How to get started right nowThe feature is available today for Pro and Max subscribers. You need the Claude desktop app installed and running on macOS. The mobile app alone won't control your computer—but you can use it to assign tasks through Dispatch, which then execute on your desktop in the background.Anthropic is framing this as a research preview specifically to understand where it falls short before a broader rollout. For a feature that has direct access to your live desktop, that measured pace is probably the right call.