Anthropic just gave its AI assistant a set of hands. Claude can now directly control macOS applications, performing mouse clicks, keyboard inputs, and screen captures across desktop apps, all while running quietly in the background so users can keep doing their own thing.
The feature, available through the Claude Desktop app for Pro and Max subscribers, represents one of the most tangible steps yet toward AI agents that don’t just talk about doing work but actually do it.
How it actually works
The new capability runs on macOS 15 (Sequoia) and later, defaulting to background mode. Rather than commandeering your mouse and keyboard, Claude operates in separate windows while you continue using your machine normally.
When traditional API connectors aren’t available for a given app, Claude falls back on screen capture and input simulation to get things done. It can read what’s on screen, figure out where to click, and type responses into fields, essentially mimicking how a human would interact with software that wasn’t built with AI integration in mind.
The system lives within two contexts: Claude Cowork, aimed at general productivity tasks, and Claude Code, focused on development workflows. Both leverage the same underlying computer control architecture, but they’re tuned for different use cases.
There’s also a mobile Dispatch feature that lets users assign tasks remotely. Start a job from your phone, and Claude handles it on your Mac.
Permissions and safety guardrails
The system requires explicit user permission before Claude takes any action during the first task in each session. Privacy prompts appear before sensitive operations, and app-level permissions give users granular control over what Claude can and can’t touch.
The initial rollout is limited to paid individual plans. Team and Enterprise subscription tiers don’t have access yet.
One honest caveat that Anthropic itself acknowledges: Claude is slower than a human at executing complex desktop tasks. The AI takes a methodical, cautious approach to each interaction, reading the screen, planning actions, and then executing them step by step.
From beta experiment to shipping product
This launch didn’t come out of nowhere. Anthropic first introduced computer control capabilities in beta form back in October 2024 as part of the Computer Use API, bundled with Claude 3.5 Sonnet. That version was more of a proof of concept, aimed primarily at developers who wanted to experiment with agentic workflows.
What this means for the AI agent race
The current limitations are real. Slower-than-human execution speed means this isn’t replacing anyone’s workflow overnight. It’s more useful for repetitive, high-volume tasks than for anything requiring real-time responsiveness. The background mode design also means Claude can’t interact with whatever you’re actively working on — it’s restricted to its own windows.
For now, the feature remains exclusive to individual paid subscribers on macOS 15 or later, with no timeline announced for Team or Enterprise availability.