From a bare OS to a private inference API.
ai-on-a-box reads what your machine 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.
- Curated configs, never inferred
- Every script readable first
- No accounts, nothing stored
Three documents, in order
profile.json
what this machine actually is — the probe reads it.
plan.json
what will run, and what will not — the catalogue derives it.
report.json
what happened, per phase, per test — the bundle writes it.
Each one is a plain file you keep. Hand the same plan back a month later and you get byte-identical scripts.
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.
Probe the machine
Read the script first — the whole thing renders on its own page. It is POSIX sh, read-only, and writes one file: what your machine actually is.
Match the catalogue
Every config is hand-written for one machine and verified on it. You get the ones that apply, the ones that apply with warnings, or — when none do — the closest and exactly which fields disagreed.
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 is downloaded.
Run the bundle
Numbered shell scripts, a checksum manifest they verify before executing, and a dry run by default. It writes a report saying what actually happened.
Machines somebody has actually done this on
The whole catalogue, and it is short on purpose. A config exists because a person stood at a machine and wrote down what happened — which is why there is no way to add your box by describing it, and why each entry says whether it was verified there or derived for the class it belongs to.
Ollama on the CPU, any Linux box with the RAM for it
buildderived-ollama-cpu-linux
1 engine verified, serving text, vision, embedding. 3 phases.
Assurance tier: derived.
Ollama in Docker, any NVIDIA Linux box
buildderived-ollama-docker-linux
4 engines verified, serving text, vision, embedding, image, speech. 6 phases.
Assurance tier: derived.
Ollama on Apple silicon, served over Metal
buildderived-ollama-macos
1 engine verified, serving text, vision, embedding. 3 phases.
Assurance tier: derived.
Ollama in Docker Desktop (WSL2), any NVIDIA Windows box
buildderived-ollama-windows
4 engines verified, serving text, vision, embedding, image, speech. 6 phases.
Assurance tier: derived.
Ollama in Docker Desktop, any Windows box with no usable card
buildderived-ollama-windows-cpu
3 engines verified, serving text, vision, embedding, speech. 5 phases.
Assurance tier: derived.
NVIDIA GX10 (GB10, unified memory)
verifygx10-gb10
1 engine verified, serving text, vision, embedding. 4 phases.
Assurance tier: derived.
Dell Precision T5820 with RTX 3090
buildt5820-rtx3090
6 engines verified, serving text, vision, embedding, image, speech. 13 phases.
Assurance tier: derived.
Three things it deliberately does not do
Each of these is a decision rather than a gap, and each one is the reason to trust the rest.
No guessing what you can run
Nothing here reads a specification sheet and decides your hardware could probably cope. Every recipe was written and run by a person; each one says whether that happened on this exact machine or on a class it belongs to, and a machine nobody has a recipe for is told so rather than handed an optimistic one.
No accounts, no database
There is no user model anywhere in either application. 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 a directory of numbered shell scripts and a checksum manifest — and `./run.sh` is a dry run unless you say otherwise.
Everything past this page needs a credential
The catalogue, the probe, the build flow and the diagnoser are all behind the gate. Your browser will ask for a username and password — the ones set by whoever deployed this box. There is no sign-up, and there is no account to create.