NS-BOS · Open-source kernel

The kernel is free. Apache 2.0.

The NS-BOS Kernel is the minimal, auditable runtime for autonomous edge AI agents — it orchestrates inference, agent lifecycle, an encrypted data plane with a hash-chained audit trail, and sandbox isolation. It runs on hardware you own and works fully air-gapped, with no cloud to log in to.

Install it Read the source

Apache 2.0 Linux Python 3.10+ No telemetry SBOM included

Read this before you install

This is a runtime, not an app. It is genuinely useful to the right person and genuinely useless to everyone else, so here is the split before any commands.

Install it if you…

run Linux — Ubuntu, Debian or Fedora
have Python 3.10 or newer (3.9 will not import it)
are comfortable in a terminal
have a model runtime and weights to point it at

Skip it if you…

are on macOS or Windows — it will refuse, by design
want a window and a chat box — that is the appliance
expect the download to contain a model: it does not
want a managed service — there is no cloud here

The kernel orchestrates inference. It ships no weights, no chat UI and no agents beyond the six runtime placeholders. You bring the model.

Install

The installer is short and does six things. Read it before you run it — that is the whole point, and it is the same posture we ask of everything else on this site.

curl -fsSL https://ngari.ai/install.sh -o install.sh
less install.sh                     # ~120 lines: read it
sha256sum install.sh                # compare with the hash below
bash install.sh

sha256 of install.sh — d2a5909a97d765f8a7025ecb268850ef772d583c2f7f1b10509bddf868f66f28

It refuses anything that is not Linux, refuses Python below 3.10 with a clear message, never calls sudo, writes only under ~/.ngari/kernel, installs the hash-pinned dependency lock, checks that the kernel imports, and then stops — it does not start the kernel and does not download a model.

Or do it by hand

git clone --depth 1 https://github.com/NGARiAI/ns-bos-kernel
cd ns-bos-kernel
python3 -m venv .venv && . .venv/bin/activate
pip install -r core/kernel/requirements.txt
NGARI_DATA_DIR=~/.ngari/kernel python -m core.kernel

Pin a version

Reproducible installs should pin a ref rather than track the branch:

./install.sh --ref v0.1.0        # a tag
./install.sh --prefix /opt/nsbos # somewhere other than ~/.ngari/kernel

The installer tracks the branch by default and prints the ref it used. If you pin --ref v0.1.0 you get exactly that tree — which predates the first-run fixes and will not start out of the box.

Verify it yourself

Everything here is checkable without asking us anything. That is the reason the kernel is open.

git clone https://github.com/NGARiAI/ns-bos-kernel && cd ns-bos-kernel
python -m pytest tests/ -q          # 32 tests
cat LICENSE                         # unmodified Apache 2.0
cat sbom/ns-bos-kernel.spdx.json    # SPDX 2.3 software bill of materials
cat THREAT_MODEL.md SECURITY.md

Hash-pinned dependencies. core/kernel/requirements.txt is generated with pip-compile --generate-hashes, so an install either matches those hashes or fails.
An SPDX SBOM and a written threat model, both in the repo.
Publisher is on the commit history. Every change is a public commit with a message that says what and why.

What is in the kernel

Inference orchestration — points at Ollama, llama.cpp or vLLM; it does not bundle them.
Agent runtime — registry, scheduler and message bus for long-lived agents.
Sovereign data plane — encryption at rest and a hash-chained audit trail.
Sandbox isolation — constrained execution for agent actions.
Network sync — mDNS discovery between your own devices, no broker.
Hardware abstraction — one codebase on aarch64 and x86_64 (Jetson, Raspberry Pi 5, generic Linux).

Questions

Is it really free? What is the catch?

Apache 2.0, no account, no key, no call-home. There is no catch in the licence. The honest commercial reason we publish it is that a sovereign runtime only matters if you can audit it — so the kernel is open and the finished appliance, the agents and the models are the product.

Does it include a model?

No. The kernel orchestrates inference; you supply the runtime and the weights. Our own models and datasets are published separately at huggingface.co/NGARiAI.

Can I run it on a laptop?

On Linux with Python 3.10+, yes — it will start. Whether it is useful depends entirely on the model you point it at.

Does it send anything to NGARi?

No. No telemetry, no analytics, no licence check. Article 5 of our constitution is "no telemetry by default" — the kernel has none at all, not even opt-in. You can verify that by reading it or by running it with the network down.

What if I want this without any of the setup?

That is what NGARi+ is: the same stack on a machine that arrives flashed, verified and ready. See the hardware.

Own the machine, or own the setup

The kernel is free and always will be. If you would rather not run a Python environment on general-purpose hardware, NGARi+ ships the same sovereign stack pre-installed and verified on NVIDIA Jetson — from $999.

See NGARi+ hardware Join AI-Corps first