Anthropic just shipped the feature that turns Claude from a chatbot into a coworker. In Claude Cowork and Claude Code, you can now enable Claude to directly control your computer — pointing, clicking, scrolling, and navigating your screen to complete tasks. When Claude doesn't have a connector to the tool it needs, it reaches for your mouse and keyboard instead. No setup, no plugins, no browser extensions.
The capability is available in research preview for Claude Pro and Max subscribers on macOS. It works especially well paired with Dispatch, which lets you assign Claude tasks from your phone and come back to the finished work on your desktop.
How It Works
Claude follows a three-tier priority system. It reaches for the most precise tool first: connectors to services like Slack or Google Calendar. When there isn't a connector, it controls your browser directly. As a last resort, it takes full screen control — scrolling, clicking to open apps, and exploring as needed. Every new application access requires your explicit permission.
This means Claude can open files, run dev tools, fill out spreadsheets, navigate web interfaces, and handle any keyboard-based task — all autonomously, but with you in the approval loop.
Dispatch: Asynchronous AI
The bigger shift might be Dispatch. Released the week prior, it creates one continuous conversation with Claude across your phone and desktop. Tell Claude to check your emails every morning from the train, pull metrics weekly, spin up a coding session for a PR, or keep a project moving according to your plan — all from your phone, while your Mac does the work.
Dispatch hit #1 on Product Hunt with 470+ upvotes within 24 hours. Commentators are calling it the moment AI shifts from "synchronous" — you watch it work — to "asynchronous" — it works while you're away. One analysis described it as "the moment AI stops being a tool you use and starts being a teammate that works."
Early Reactions: Promising but Raw
The buzz is significant but tempered. CNBC and Engadget covered the announcement as a milestone in the AI agent race, noting it comes as companies compete to build autonomous agents following OpenClaw's viral success. MacStories' hands-on review rated reliability at "about 50/50" — file retrieval, database queries, and summarization tasks work well, but opening applications, sending messages through Slack, and terminal access remain unreliable.
The consensus: Dispatch isn't production-ready when you're away from your desk, but it represents a meaningful step toward the async AI workflow everyone is building toward. The feature requires your Mac to stay awake with the Claude app running and only supports single-threaded execution — one task at a time.
Safety Guardrails
- Prompt injection detection — automatic activation scanning within the model to detect malicious inputs
- Explicit permission required before accessing new applications
- User kill switch — stop Claude at any point during execution
- App allowlisting — some applications are off-limits by default
- Sensitive data advisory — Anthropic recommends starting with trusted apps and avoiding sensitive data
Computer use is still early compared to Claude's text and coding abilities, and Anthropic is transparent about that. But the direction is clear: the AI agent that can read your screen, control your apps, and take assignments from your pocket is no longer theoretical. It's in research preview, it's rough around the edges, and it's exactly where every AI company is heading.