I Built a Private ChatGPT for My Family on a Mac Mini

For about a year, my household had three separate AI subscriptions. My wife had ChatGPT Plus for work emails and meal planning, I had my own stack, and my oldest kid kept borrowing accounts for homework help. That’s roughly $40–60 a month, every prompt shipped to a third-party server, and zero visibility into what the kids were actually asking.

So I repurposed the M4 Mac Mini that was already running as a Plex and Time Machine box, and in one evening turned it into a private ChatGPT for the whole family. Everyone in the house — laptops, iPhones, the iPad in the kitchen — now talks to the same local AI through a normal browser. Nothing leaves our network. Here’s the exact build.

The hardware and why a Mac Mini is the sweet spot

The base M4 Mac Mini with 16 GB unified memory is enough for a genuinely useful family model. With 24 GB or 32 GB you get headroom for larger models and multiple simultaneous chats. The key Apple Silicon advantage: unified memory means the GPU can use almost all of the RAM for the model, and the whole machine idles at around 4–5 W. Mine, measured at the wall with a TP-Link Kasa smart plug, draws 3.8 W idle and peaks at about 38 W while generating tokens. Even if the family hammered it two hours a day, that’s under €2 a month in electricity where I live. Compare that to $20/month per ChatGPT Plus seat and the Mini pays for itself in well under two years — faster if you already own one.

Step 1: Ollama as the engine

Install Ollama and pull a model that fits your RAM:

brew install ollama
brew services start ollama
ollama pull llama3.1:8b
ollama pull qwen2.5:14b

On 16 GB, llama3.1:8b (Q4, ~4.9 GB) is the workhorse — I get around 28 tokens/sec on the M4, which feels indistinguishable from ChatGPT in a chat window. With 24 GB+ I’d make qwen2.5:14b the default; it’s noticeably better at homework-grade math and multilingual answers (we’re a Czech/English household, and Qwen handles Czech far better than Llama).

One non-obvious step: by default Ollama only listens on localhost. To serve the LAN, it needs to bind to all interfaces:

launchctl setenv OLLAMA_HOST "0.0.0.0:11434"
brew services restart ollama

Also go to System Settings → Energy → enable “Start up automatically after a power failure” and set the Mac to never sleep (Displays can sleep; the system can’t). An AI server that naps is useless.

Step 2: Open WebUI for the ChatGPT experience

Ollama alone is a terminal tool. Open WebUI is what makes this feel like a product your family will actually use. I run it via Docker:

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

Now anyone on the network opens http://macmini.local:3000 and sees a clean ChatGPT-style interface. The first account you create becomes the admin. Then, in Admin Panel → Settings → General, set new signups to “pending” so the kids can register themselves but you approve each account.

This multi-user part is what surprised me most. Each family member gets their own login, their own chat history, their own system prompt. For my 11-year-old I set a default model with a system prompt along the lines of “You are a patient tutor. Explain step by step, never just give the final answer to homework problems.” In Admin Panel → Users you can restrict which models each user sees — the kids get the tutor preset, my wife gets the unrestricted 14B.

Step 3: iPhones, the kitchen iPad, and the PWA trick

Nobody in my family wants to type a URL. Open WebUI is a proper progressive web app, so on each iPhone: open Safari → http://macmini.local:3000 → Share button → “Add to Home Screen.” You get an icon that launches full-screen, no Safari chrome, indistinguishable from a native app. My wife genuinely did not realize for two weeks that “her AI app” was running on the Mac in the office closet.

For access outside the house, I added Tailscale — install it on the Mini and every phone, sign in with the same account, done. It builds a WireGuard mesh, so http://macmini.your-tailnet.ts.net:3000 works from anywhere with zero open ports on the router, zero port forwarding, zero exposed services. The free plan covers up to 100 devices, which is more than enough for a family. My daughter has used the homework tutor from a train with no idea anything clever was happening.

The parental angle nobody talks about

This is the part that turned my skepticism into conviction. When a kid uses ChatGPT, their prompts go to OpenAI under an account you may not control, with retention policies you can’t audit. On Open WebUI, as admin you can see usage per account, and every conversation lives in a SQLite database on a machine you own. I told the kids upfront: this is house AI, and house rules apply — same as the family computer in 1998.

The local model also can’t browse, can’t generate images of real people, and answers from frozen weights — which, for a child’s homework helper, is honestly a feature. No rabbit holes, no upsells, no “continue in the app.” And when my son asked it something embarrassing about puberty, that question stayed on our hardware instead of becoming a data point in someone’s ad profile. That alone justified the project for me.

Costs, limits, and honest caveats

What this setup does brilliantly: everyday Q&A, writing help, translations, homework tutoring, recipe math, summarizing pasted text. With Open WebUI’s built-in document upload (it does RAG out of the box), my wife dumps PDFs of school newsletters in and asks “what do I actually need to do this week?”

What it doesn’t do: an 8B model is not GPT-5.2. It will lose to frontier models on complex reasoning, current events, and long multi-step tasks. My rule of thumb — the local server handles 80% of family queries, and I keep a single paid frontier subscription for my own deep work. We went from three subscriptions to one.

Total damage: one evening of setup, ~€20/year of electricity, and hardware many Mac households already own. If you’re buying new for this purpose, get the 24 GB Mini — the jump from 8B to 14B models is the single biggest quality upgrade per dollar in local AI right now.

The unexpected payoff is cultural. AI in our house stopped being a mysterious cloud thing and became an appliance, like the router. The kids understand that a model is a file on a computer, that it can be wrong, and that Dad can read the logs. That’s a better AI education than any school curriculum I’ve seen — and it started with a Mac Mini and two terminal commands.