Anthropic just introduced native “computer use” inside the Claude desktop app, and it’s not the usual browser-control agent. This is full-on Mac OS control: Claude can take screenshots, move your mouse, type on your keyboard, and operate basically any app you allow—per app, per session—through accessibility permissions. In the video, I break down what it actually does, how it fits with features like Dispatch (including using your computer remotely from your phone), and why this is a pretty big step toward automating real knowledge work rather than shipping gimmicky demos. I also go through the practical limitations and the stuff you should be cautious about before you hand over the keys to the kingdom. Your desktop needs to be active, complex tasks may take multiple tries, and it can be painfully slow because you’re watching it “human-click” through UI. On the safety side, there are explicit constraints: it won’t do things like banking logins, stock trading, or entering sensitive credentials. And the bigger meta-point: with memory plus screenshots, this starts to look like a supercharged version of the “Recall” concept—so you need to be thoughtful about what you expose. I demo it opening DaVinci Resolve and finding my latest project, then test a banking scenario to show the guardrails in action.