The Seed, Built on a Bench

When we put Node0 in the open on 20260922, the Seed was a design: the smallest version of Node0 anyone can start from, as a ladder of rungs, with rung 0 in progress. Between 20260923 and 20260928 an agent built rungs 0 to 4b on a bench of spare hardware, working from those pages, with Christoph doing the hands: plugging in, pressing buttons, deciding. Everything it built, measured and got wrong is now published, at oznog.com/node0/seed and in the build’s repository .

What is there #

A page for every rung, generated from the repository so a fact is written once. Each says in plain terms what the rung gives you, what to buy, what it costs, what still fails, the decisions with what would change them, and the proof. Rungs 0 to 4b ran on the bench: the laptop with backups, the infrastructure box, the agent’s own box, the core box, and a compute box serving four kinds of models behind one gateway. Rungs 5 and 6, the edge with a second site and a dedicated firewall, are designed, and the three wifi networks of rung 5 are already built and tested.

The configuration exactly as it ran, in the repository: the site’s settings, the scripts, the monitoring rules with their tests, the measurements and the pitfalls, with secrets and serials removed. It is what a builder’s agent reads to repeat the bench, and what a person reads to see how a piece looked when it worked.

The Seed page keeps the whole ladder in view: a status table that says which rungs have actually been run, a build log, the costs by rung, and the projects the Seed runs on. Through rung 3, a complete site without local AI, the hardware is $2,060 lean and $3,780 full, on prices checked 20260920 and partly rechecked 20260928. On the bench, rungs 1 to 4 behind the UPS drew about 49 W at rest; the router and switch were not measured.

A page for people who are not technical, why run your own, and when not to . Apple, Google and the AI companies already do most of this for you. That page sets out what they really cost, which is mostly not money but what happens to your data, what running your own costs instead, and when a hosted service is the better answer, piece by piece.

What the bench proved #

A claim that something recovers is a belief until the failure has been caused. So the bench caused them.

• With the infrastructure box shut down, we cut its power and restored it, touched nothing, and it started by itself with every service answering 1 minute 39 seconds after the power returned (20260924).
• We rebuilt the infrastructure box on a spare mini PC from the recovery pack alone, in 1 hour 48 minutes, 70 of them spent on traps now written down (20260928).
• A deploy that took the agent box’s network away rolled itself back to the last good version in 22.5 minutes, with nobody at the box (20260928).
• With the compute box’s model server stopped, the private routes returned errors and sent nothing out; only the coding route answered, from the hosted model, and the gateway metered it (20260928).
• The guest and IoT wifi networks reached the internet and were refused everything else, in 93 checks with none failed (20260928).

What broke #

The mistakes are left in, because they are the part a builder is least able to find alone. The array does not start by itself after a reboot unless you turn that on, and on 20260924 that left the whole site down until someone started it by hand; the only alarm was the external dead-man’s switch. A power cut sent no “on battery” alert, because the box began shutting down before the one-minute poll had run. A 10G link went up and down 125 times in 17 minutes after a cold start. The infrastructure box’s power button sits just above its USB ports, and pulling a stick from a running box pressed it. Each is written up in the pages, with the check that now exists.

Why it matters #

Money is rarely the reason to run your own. The reason is what happens to your data: what it is used for, and now what AI is trained on, for you or against you. We choose to be the customer, the one a service answers to, rather than the product. That is connected sovereignty: not cutting yourself off, but holding what must stay private and choosing, deliberately, what to use from others. The Seed is that choice at the smallest size, and the page on why to run your own also says when not to.

Who it is for #

It is written to be read by a person and used by their agents, and it gets technical quickly. That detail is for the agent. You do not need to follow all of it: ask your agent what a thing means, why you would want it, and what happens if you skip it. It has read the same pages.

What comes next #

The rest of rung 5 and rung 6, built on the bench the same way. Runbooks for each rung, written by a fresh agent building from the published pages alone, so each one is tested by being followed. And the part we expect to matter most: an agent that looks after the site, reading every alert, checking that backups and restore tests keep passing, applying updates through the repository, and fixing routine problems within permissions you write down. Node0 already runs one.

What we ask #

Build one, at whatever rung you need, and tell us what the pages got wrong or left out, as an issue on the repository or by message . The content is CC BY 4.0 and the code is MIT, so take what is useful and keep the notice.