A private ChatGPT for a family of four on a Mac mini

What they built

Jakub Jirák's household had three AI subscriptions, roughly $40-60 a month, with every prompt shipped to a third party and no visibility into what the kids asked.

In one evening he turned the M4 Mac mini already running Plex and Time Machine into a private ChatGPT for laptops, iPhones and the kitchen iPad. Each family member has their own login, history and system prompt.

Setup

• M4 Mac mini, 16 GB
• Ollama with llama3.1:8b (Q4, ~4.9 GB) as workhorse; qwen2.5:14b recommended with 24 GB+
• Open WebUI in Docker on port 3000
• OLLAMA_HOST bound to 0.0.0.0 so the LAN can reach it

How it works

1. Ollama serves models on the LAN
2. Open WebUI provides accounts; new signups are set to 'pending' for admin approval
3. Per-user model restrictions: kids get a tutor preset, adults the unrestricted 14B
4. Chats live in a SQLite database on the Mac

Results & numbers

• ~28 tokens/sec for llama3.1:8b on the M4
• Qwen handles Czech far better than Llama for this bilingual household

Gotchas

• Ollama only listens on localhost by default
• The Mac must never sleep and should auto-restart after a power failure

“Explain step by step, never just give the final answer to homework problems.”— Tutor system prompt, Jakub Jirák

Do it yourself

1. brew install ollama && brew services start ollama; pull llama3.1:8b
2. launchctl setenv OLLAMA_HOST "0.0.0.0:11434" and restart Ollama
3. Run Open WebUI in Docker (command below)
4. Create the admin account first, then set signups to pending
5. Give kids a tutor system prompt and restrict which models they see

docker run -d -p 3000:8080 \
  -e OLLAMA_BASE_URL=http://host.docker.internal:11434 \
  -v open-webui:/app/backend/data \
  --name open-webui --restart always \
  ghcr.io/open-webui/open-webui:main