All articles

Where do you find the two hours?

Setting up your own AI at home takes a lot of steps. Here is what we do before The Box ships, and what is still up to you.

Most people who want private AI stop at step one. Even people who know a lot about it.

Mike, one of our two co-founders, used AI mostly to write code. He and Asha, the other co-founder, set one up for themselves first. It did work. But each new job meant more work: which of his files would it need, and how would he hook them in?

"If it's going to take 2 hours, that's 2 hours away from developing," he told us. "Where do I find those 2 hours?"

So for most of that work, he just did not use it. Or he used an AI on the web that he would rather not have given it to.

On our calls, people tell us again and again that setting it up is a big wall. One person on our waitlist said they "had the idea but not time to do the tinkering."

What doing it yourself takes

The tools and the AI are free to get. The hard part is six choices, where each one can change the rest.

1. Pick the machine. The AI has to fit in its memory, with room left for you to talk to it. In some small machines, the memory is stuck to the board. You can't add more later.

2. Pick the AI. One free AI has 21 kinds on its page, from under 10 GB to over 35 GB [1]. Small ones are fast. Big ones do better on hard work. Which one will fit your machine?

3. Pick the tool that runs it. A well-known one is llama.cpp. It comes in a build for each kind of chip, like Nvidia, AMD and Apple [2]. You have to pick the one that fits yours.

4. Get a place to talk to it. Some of these tools have a plain chat page [2]. For more, most people add one more tool. One, Open WebUI, tells you to start it by typing this. And that is after you add two more tools [3]:

docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:main

If that line means nothing to you, that is the point. If it does, you know it is never just one line.

5. Lock it. It is only yours if no one else can get to it. Ollama, one more tool that runs AI, at first lets in only your own machine [4]. To use it from your phone, you have to change that. Get it wrong, and people you don't know may be able to use it too. A team at SentinelOne and Censys looked for about ten months. In January 2026, they said they had found more than 175,000 machines that ran Ollama and were open to the whole web [5]. Many had no lock at all.

6. Keep it up to date. Each part gets new versions on its own time. On Linux, the Ollama help page says to run its install again [4]. Each change can make something stop working.

Want to use it away from home? That is a seventh job. Only then do you get to the part Mike was stuck on: making it work with your own stuff.

What we do before it ships

The Box is a small Framework Desktop with an AMD Ryzen AI Max+ 395 chip. It has 64 GB or 128 GB of memory. We chose it so the AI we ship will fit, with room for long talks.

Before it leaves us, we do steps 1 to 5. We made the tool that does this, called Halo. The rest are free, open projects made by other people. We name each one on purpose:

  • The AI: Qwen 3.6 and Gemma 4. Anyone may get them and use them for free.
  • The tool that runs it: llama.cpp.
  • The helper: Hermes Agent, from Nous Research. It lets the AI do tasks, not just talk.
  • The chat page: Hermes WebUI.
  • The memory: Mnemosyne. It keeps what the AI knows about you on the box.
  • The lock: a password you choose, so only the people you let in can use the chat.

What setting it up looks like for you

You need the box, a plug, your home Wi-Fi name and password, and a phone.

Plug it in. Give it about a minute to start.

Join the box's own Wi-Fi. It is called box-setup.

Scan the code on the label. Or type http://box.local into your web browser.

Sign in, then pick your password. Use the one on the label first. Then make your own, with twelve or more letters, numbers or marks. A short phrase will do.

Link the box to your home Wi-Fi.

Start to talk to it. Switch your phone back to your home Wi-Fi. Open http://box.local again. The chat is there.

Two extras can wait. Tailscale is a tool that lets you use the box when you are away from home. You make your own account. It is free for up to six people [6]. Signal lets the box send you a note on your phone.

That is it. You don't type code.

Help, in writing

On a call, one person asked if "plug and play" meant they would not have to think about how it works. Mostly, yes. And when you do, you are not on your own. Our Terms of Sale, the rules of the sale, say what you get:

  • At least the first two weeks: help to get you going. Book a time, 9 to 5 Eastern, Monday to Friday. We can fix what is wrong, set up the AI, and hook up your first tasks and tools.
  • The first twelve months: if something is wrong with the box, we help fix it for free.
  • Other help after the first two weeks, like new tasks or more AI set-up, is up to us. It may be free, or it may cost money, unless your bill says so.

If we need to look at your box, you let us in. In Settings, the box shows a code with 6 numbers. You read it to us. Then we can get in for up to four hours. While we are in, we can see your box. A yellow-orange bar stays on your screen the whole time. You can end it when you like.

What's still up to you

  • Which size. The memory is stuck to the board, so your choice is for good. The 128 GB box runs bigger AI, and more people can use it at once. You can send it back within 30 days. But we set up each box before it ships, so we keep 15% of the price, even if you never took it out of the box. You also pay to ship it back. Not sure? Ask us first.
  • Your password. It is the lock, so keep it safe. If you lose it, you can make a new one on a phone or computer that is still signed in. If none is, the only way in is to wipe the box and start over. That wipes out all your stuff on it. We will walk you through it.
  • Saying yes to new versions. Once a day, the box asks us if there is a new one. It only tells us which one it has. The service that sends out new versions also sees where the ask came from. It keeps that for 30 days, only to fix problems. You can turn this off. Nothing new goes on until you say yes. If a new one does not pass the box's own checks, the box goes back to the old one. We do not say we will make new ones, or when.
  • What it is good at. It is good at the writing, reading and tasks of most days. AI that runs at home may not do the very hardest work as well. That is the trade: your words stay with you. Try the demo on our home page and see.
  • What to use it for. The Box takes care of the steps above. In the first two weeks, we help you hook up your own stuff. After that, new work is mostly yours to set up, though you can ask us for help. Picking that work is the part worth your two hours.

If this sounds like you

If private AI has been on your to-do list, The Box can take it off. You pay one price for the box, the set-up, and a year of free help if the box has a problem. You don't pay each month.

We build boxes in groups. When a group is open, you can buy one on our pricing page. When it is sold out, join the waitlist. We will tell you when the next one opens.

See pricing
Why private means a box on your desk
What owning your AI means
The case for paying once

Sources

Public sources (the only ones to publish with the page). Sources 1, 2, 5 and 6 were re-opened on 2026-09-28. Sources 3 and 4 were checked in the previous pass, and the reviewer re-verified 4; neither was re-opened in this pass.

  1. One 27B model (Qwen3.6-27B) offered in 21 quantized GGUF versions, 9.39 GB to 35.3 GB (as of September 2026; live repo, count can change) — https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/tree/main
  2. llama.cpp: CUDA (Nvidia), HIP (AMD), Metal (Apple), Vulkan and other backends; built-in web UI ("Built-in web UI against llama serve") — https://github.com/ggml-org/llama.cpp
  3. Open WebUI Docker start command (quoted exactly from the README, "if Ollama is on your computer") — https://github.com/open-webui/open-webui
  4. Ollama binds 127.0.0.1:11434 by default; changed with OLLAMA_HOST; on Linux, update by re-running the install script — https://docs.ollama.com/faq
  5. SentinelLABS and Censys (Bernadett-Shapiro and Cutler, 29 January 2026): 175,108 unique Ollama hosts in 130 countries over 293 days of scanning, "internet-reachable," "accessed without authentication, usage monitoring, or billing controls" — https://www.sentinelone.com/labs/silent-brothers-ollama-hosts-form-anonymous-ai-network-beyond-platform-guardrails/
  6. Tailscale Personal plan: free, up to 6 users (checked September 2026) — https://tailscale.com/pricing