Put local inference on the idle laptop, not the server
Use the laptop's chip for free speech and small AI models so the home server stays responsive.
how it works
Kokoro TTS on MLX and a small Ollama model on the Air, reached over the tailnet with fallbacks when it sleeps; caffeinate LaunchAgent keeps it reachable on AC.
Technique: A pattern or hard-won lesson, written down so it isn't learned twice.
The code for this one lives in a private repo. This page explains the idea so you can build your own.