Flagship product

ai-on-a-box — from a bare OS to a private inference API

Provisioning for self-hosted AI appliances. It reads what your machine actually is, works out what can run on it, and hands you the scripts that get it there — every one of them readable before anything executes.

What it is

Not an installer — a provisioner you can audit

Standing up local AI on your own hardware is rarely one command. It is drivers, runtimes, disks, engines, and a dozen decisions that only make sense on that box. ai-on-a-box turns that into three plain files you keep: a profile of what the machine is, a plan of what will and will not run on it, and a report of what actually happened. In between you get numbered shell scripts and a checksum manifest — a dry run by default, and readable in full before a single line executes.

0
Plain files you keep — profile, plan, report
Hand-verified
Every config written at the machine itself
0
Accounts, databases, records that you were here
How it works

Four steps, and you can stop after any of them

Nothing is stored, so coming back tomorrow costs you nothing but the file you kept.

1

Probe the machine

Read the script first — the whole thing renders on its own page. POSIX sh, read-only, and it writes exactly one file: what your machine actually is.

2

Match the catalogue

You get the configs that apply, the ones that apply with warnings, or — when none do — the closest one and exactly which fields disagreed.

3

Read the plan

What will run, in order. What will not, and why. Which disk gets claimed, which engines start on their own. In sentences, one screen before anything downloads.

4

Run the bundle

Numbered shell scripts and a checksum manifest they verify before executing — a dry run unless you say otherwise. It writes a report of what actually happened.

Three documents, in order

Files you keep, not state we hold

Hand the same plan back a month later and you get byte-identical scripts.

profile.json

What this machine actually is — CPU, memory, GPU, disks, OS. The probe reads it, and it is yours: nothing is uploaded on your behalf.

plan.json

What will run, and what will not — derived from the catalogue and explained in sentences. The same plan always produces the same bundle.

report.json

What happened, per phase and per test — written by the bundle on the machine, so a run that half-succeeded says so instead of looking clean.

The catalogue

Machines somebody has actually done this on

It is short on purpose. A config exists because a person stood at that machine and wrote down what happened.

Dell Precision T5820 + RTX 3090

4 engines verified — text, vision, embedding and image — across 11 phases.

NVIDIA GX10 (GB10, unified memory)

Verified across 4 phases — serving text, vision and embedding.

Ollama in Docker, on NVIDIA

Two phases on any NVIDIA Linux box — plus a Docker Desktop / WSL2 config for Windows.

Apple silicon, served over Metal

And for everything else, Ollama on the CPU — any Linux box with the RAM for it.

Why it holds up

Three things it deliberately does not do

Each one is a decision rather than a gap — and each one is the reason to trust the rest.

No rules engine

Nothing here infers what your hardware could probably run. If nobody verified an engine on a machine, that machine's config does not list it — so the catalogue is short, and everything in it is true.

No accounts, no database

There is no user model anywhere in it. Nothing is kept between requests: no profile you uploaded, no plan you derived, no record that you were here.

No black box

The probe renders in full before you download it. The plan is explained in sentences before the bundle exists. The bundle is numbered shell scripts and a checksum manifest — and the run is a dry run unless you say otherwise.

Want AI running on hardware you own?

ai-on-a-box is a SunFox product — and the same team will size the machine, provision it, and stand up the private inference API behind your firewall.