Mac Studio as an always-on, LLM-driven smart-home brain

What they built

The author runs Home Assistant, an Ollama model and a fully local voice pipeline on one Mac Studio. The LLM turns vague human requests into actual Home Assistant service calls.

Examples from the house: 'get the house ready for the morning' combines weather, calendar and presence; 'it's stuffy in here' checks which room the phone is in and the CO₂ sensor, then opens a window. One sentence about overnight guests reconfigures fourteen things.

Setup

• Mac Studio (latency measured on an M2 Ultra)
• Home Assistant in Docker
• Ollama with qwen3:14b
• wyoming-whisper (small model) and Piper
• ESP32 voice satellites in three rooms, about €15 each

How it works

1. Home Assistant's Ollama integration exposes devices to the model as tools
2. The model emits real service calls such as light.turn_on and cover.set_position
3. Assist pipeline: Whisper → LLM → Piper, all on the LAN

Results & numbers

• Simple device command: 1.8-2.5 s end to end (Whisper ~0.5 s, LLM ~1-1.5 s)
• Multi-step contextual command: 3-5 s; Siri/Alexa are ~1-1.5 s on simple commands
• Idle power 9-11 W, 90-120 W during inference; about 0.3 kWh a day (~€0.08)

Gotchas

• On a base M-chip Mac mini with an 8B model, multi-step commands stretch past 6 seconds

“the LLM isn't "smarter" than automations — it's a compiler from human vagueness to the same service calls.”— Think Different

Do it yourself

1. Run Home Assistant in Docker with host networking
2. Install Ollama and pull qwen3:14b
3. Add the Ollama integration and expose only the devices you want controlled
4. Build a local Assist pipeline with wyoming-whisper and Piper
5. Add cheap ESP32 voice satellites per room

docker run -d --name homeassistant --restart=unless-stopped \
  -v ~/homeassistant:/config --network=host \
  ghcr.io/home-assistant/home-assistant:stable