A private AI home server for your whole family

The endgame of local AI isn’t a chat window on your laptop — it’s a household utility. One machine serves phones, laptops and the TV, with per-person accounts and history, entirely inside your walls.

Hardware: quiet beats fast

A used mini-PC or small tower with a 12–24GB GPU covers most families: 8B models for everyone, 32B for power users. Prioritize noise (case fans, GPU zero-RPM idle) over peak speed — this thing lives in your hallway, not a datacenter.

The stack

Three containers, one docker-compose file:

services:
  ollama:
    image: ollama/ollama
    volumes: ['ollama:/root/.ollama']
    restart: unless-stopped
  open-webui:
    image: ghcr.io/open-webui/open-webui:main
    ports: ['3000:8080']
    environment: { OLLAMA_BASE_URL: http://ollama:11434 }
    volumes: ['open-webui:/app/backend/data']
    restart: unless-stopped
  whisper:
    image: onerahmet/openai-whisper-asr-webservice:latest
    ports: ['9000:9000']
    restart: unless-stopped
volumes: { ollama: {}, open-webui: {} }

• Ollama runs models and exposes the API.
• Open WebUI adds accounts (so your kid’s history isn’t yours), RAG uploads and model switching.
• Whisper gives the household a transcription endpoint.

Access from anywhere (that you control)

Two sane options, both free:

1. Tailscale — a private mesh network. Install on family devices, reach http://server:3000 from school or the office. Zero exposed ports.
2. LAN-only — honestly, most families only use it at home. Done.

Do not port-forward the UI to the internet without serious hardening.

Model strategy for shared hardware

• Keep one fast default (Qwen3 8B) loaded for instant responses.
• Schedule big-model time: 32B reasoning at night is fine on one GPU.
• Open WebUI lets you set per-user model permissions — give the TV a small model.

Maintenance is one cron line

0 4 * * 1 docker compose pull && docker compose up -d

Weekly updates, zero thought.

Where to next

• Tier guide for hardware picks
• ROI calculator — a family sharing one rig is where the math gets absurd