Anthropic just gave Claude the ability to actually use your Mac — click things, open apps, navigate your browser. I’ve had a few days with it. Here’s what’s genuinely useful vs. what’s still rough.
What it actually is
Claude sees your screen and can click, type, scroll, and open apps. That’s the whole thing. The design decision I find interesting: it tries connectors first. If you have Slack or Google Calendar linked, it uses those APIs directly. Mouse and keyboard are the fallback when no connector exists.
That fallback is exactly why this matters. I’ve been building structured MCP workflows for months — Slack reads, Linear ticket creation, the works. Computer use isn’t a replacement for those. It’s what you reach for when no connector exists.
The Dispatch workflow
This is the piece that changes my daily routine. Dispatch lets you send a task from your phone, Claude works on your Mac, and results are waiting when you return.
“Check the Vercel deploy, Slack me the URL if it’s green.” I send that before walking into a meeting. By the time I’m back at my desk, it’s done.
I’ve been building something similar manually — a cron-based Claude assistant I call Travis, which I covered in how I actually use Claude Code. Dispatch does a version of this without the glue code.
To enable it: Desktop app → Settings → Computer Use. Keep the app running and awake.
Use cases I’d actually reach for
Deploy check without a script — “Open Safari, go to vercel.com, find the latest deploy for my-project, Slack the URL to #dev.” I could write an MCP connector for this. I haven’t. Now I don’t have to.
Async Figma review — “Open Figma, screenshot the Button component, describe what’s changed.” My design team moves faster than I check Figma. This gets me a diff without switching apps.
One-off admin tasks — anything too niche to justify a script but too annoying to do manually.
Where it still struggles
Complex multi-step workflows still derail mid-task more often than not. It’s noticeably slower than a targeted MCP connector — screen interaction has overhead that an API call doesn’t. And since it’s UI-based, any layout change can break a workflow.
The mental model that makes it click
MCP connectors are still the right tool for structured, repeatable tasks. As the shift toward agent-first software shows, proper integrations give Claude clean, token-efficient access to data. Computer use is the fallback layer — slower, less reliable, but more flexible because it works anywhere a human can click.
Think of it as: connector = precision tool, computer use = duct tape. Both useful. Not interchangeable.
It’s not magic yet. But it’s the first time I’ve watched Claude do something I couldn’t have automated myself without writing a full script.
