Tap ⌥Space, say “tidy my desktop, play some jazz, order a pizza and plan my day”, and Alfred-OS runs all four at once: a small, fast model decides which agent takes each task and whether it is destructive, then Claude agents do the work on the Mac and in a real browser.

I built everything between your voice and the orchestrator. The overlay records you, speech-to-text turns it into text in about 0.6 s, and the task API answers at once while the work runs in the background. The overlay says “On it”, shows a toast when a task is done, and keeps an agent’s question on screen until you answer it with a button or a spoken yes or no. A tray window lists every task with its live steps. In a second PR I moved the browser agent onto Claude Opus 5, routed note requests to the desktop agent, and wrote the demo guide.

The demo reel has music and no narration; everything it says is on screen.