Anthropic new feature gives Claude control of its users’ mouse, keyboard and screen

In a major expansion of its “Agentic” ecosystem, Anthropic has officially released a new “Computer Use” research preview that allows Claude to take direct control of a user’s mouse, keyboard, and screen.

While a developer-only version of this technology first appeared in late 2024, the version launched yesterday (March 23, 2026) is a far more advanced “human-like” interface. It is specifically designed to work with Claude Cowork and Claude Dispatch, allowing users to assign complex tasks from their phones that Claude then executes autonomously on their Mac.

1. How It Works: “The Keys to the Mac”

Unlike traditional software that uses invisible APIs to talk to other apps, Claude now “sees” the computer screen exactly as a human does.

• Visual Perception: Claude takes frequent screenshots of your desktop and uses a “pixel-counting” technique to identify where buttons, text fields, and menus are located.
• Autonomous Input: The model can move the cursor, click, scroll, and type to navigate through any application—even those without an official AI integration.
• “Dispatch” Integration: Using the new Claude Dispatch mobile tool, a user can be away from their desk and send a prompt like: “Fill out that 50-row expense spreadsheet using the receipts in my ‘Downloads’ folder and then email it to finance.” Claude will wake the computer and perform the task.

2. Availability & Requirements

The feature is currently in a Research Preview with several strict limitations for safety and performance:

• Platform: Currently exclusive to macOS (Windows and Linux support is expected later in 2026).
• Access: Restricted to paid Claude Pro, Max, and Claude Code subscribers.
• Setup: Requires the latest versions of both the Claude Desktop and mobile apps to be paired and updated.
• Connectivity: Works best when paired with OpenClaw, an open-source ecosystem that helps AI models interact with third-party software.

3. Safety and “Permission-First” Design

Giving an AI control over a mouse and keyboard is a significant security risk, which Anthropic is addressing through a “human-in-the-loop” approach:

1. App-Level Permissions: Claude must request explicit permission before it can interact with a new application (e.g., “Claude wants to access Slack. Allow?”).
2. The “Kill Switch”: Users can monitor Claude’s actions in real-time and hit a “Stop” button at any moment to seize back control.
3. Restricted Data: Anthropic strongly advises against giving Claude access to sensitive financial or personal data during this research phase, as it remains vulnerable to “prompt injection” attacks.

The Evolution of “Computer Use”

“The future where I never have to open up my laptop to get work done is becoming real very fast,” noted Alex Albert, a researcher at Anthropic, following the launch.