Anthropic just shipped the feature everyone saw coming but nobody expected this soon: Claude can now take over your Mac and do your work while you go grab coffee.
They're calling it Claude Desktop Agent, and it's available now for Claude Pro and Max subscribers.
What It Actually Does
You message Claude a task — from your phone, your browser, wherever — and Claude opens apps on your Mac, navigates your browser, fills in spreadsheets, moves files around. Anything you'd do sitting at your desk, it does for you.
Anthropic demoed a scenario where someone is running late for a meeting. They text Claude: "export my pitch deck as PDF and attach it to the meeting invite." Claude opens Keynote, exports the file, finds the calendar event, and attaches it. No hand-holding required.
This works through Dispatch, a feature inside Claude Cowork that gives you a persistent conversation thread with Claude across devices. You start a task on your phone during your commute, Claude picks it up on your Mac at home.
How This Is Different from Computer Use
Anthropic has had "computer use" in research preview since late 2024. That was Claude looking at screenshots and clicking things — functional but clunky. The desktop agent is a step beyond. It has deeper OS integration on macOS, meaning it can interact with native apps more reliably rather than just pixel-hunting across screenshots.
The key difference: previous computer use felt like watching someone remote-desktop into your machine while wearing oven mitts. Desktop Agent feels more like handing your laptop to a competent intern.
The OpenClaw Connection
Here's the part Anthropic won't love hearing: this launch is clearly a response to OpenClaw going viral. CNBC's own reporting says Anthropic wants "an AI agent that can rival the viral OpenClaw." Nvidia's Jensen Huang called OpenClaw "the next ChatGPT" last week.
OpenClaw already does what Claude Desktop Agent does — and it works with multiple AI providers, not just one. It runs locally, connects through WhatsApp or Telegram, and controls your machine. The open-source project has been doing this for months.
That said, Anthropic has one big advantage: integration depth. When you control the model AND the client, you can optimize the experience in ways a model-agnostic tool can't. Claude knows its own capabilities. OpenClaw has to work around the edges of whatever model it's talking to.
The Safety Question
Anthropic — being Anthropic — buried a warning in their announcement: "Computer use is still early compared to Claude's ability to code or interact with text. Claude can make mistakes."
They built in permission gates. Claude asks before accessing new apps. It won't go rogue and start emailing your contacts or deleting files without explicit approval. Smart move, especially given how fast this space is moving.
But the real risk isn't Claude going haywire. It's people over-trusting it. If Claude fills in a spreadsheet wrong or sends the wrong attachment, you might not catch it until your boss does. The "set it and forget it" pitch is compelling, but autonomy without oversight is how mistakes compound.
Who Should Use This
If you're already paying for Claude Pro ($20/mo) or Claude Max ($100/mo), this is worth trying. Best use cases:
- Repetitive admin tasks: filing expenses, organizing files, updating docs
- Multi-app workflows: anything that requires copying data between apps
- Meeting prep: gathering docs, creating agendas, sending reminders
Worst use cases: anything involving sensitive data you wouldn't show a new employee. Claude is smart, but it's not your employee. It doesn't understand context the way you do.
My Take
This is Anthropic's most aggressive product move yet. They've been the "safety-first" AI lab, the ones who publish responsible scaling policies while OpenAI ships features. Launching a full desktop agent — even with guardrails — signals they're done playing defense.
The agent era is officially here. The question isn't whether AI will control your computer. It's whether you'll trust it to.
Top comments (0)