Show HN: HART OS – an open-source AI OS built so frontier AI needs no datacenter

Hevolve Hive Agentic Runtime

Democratic frontier intelligence with zero lock-in, fronted by an agentic OS.

Runs on your hardware. No vendor silicon, no subscription, no broker. Local-first, federated, OpenAI-compatible.

An assistant that runs on your own machine. No subscription. Works with the wifi off. 8GB of RAM is enough. What you type stays on the device because there is nowhere else for it to go, and you can watch the network to check.

On a hard question a frontier model beats anything that fits on a laptop. Most of what people ask in a day is not that. This is for the rest.

Ready today: Nunba for Windows, Linux and Android. This repo is the runtime underneath it.

Training needs tight interconnect, so training clusters are dense. That part is physics and a reader here was right to pull me up on wording that implied otherwise.

Interconnect binds one kind of training: synchronous, large batch, everything

in step. That is not the only kind. Learning here is incremental and gossiped,

built up on consumer hardware as nodes use it, with federated_aggregator.py

doing periodic aggregation rather than a tight all-reduce. A world model

accumulated that way does not need everyone in one building, because nobody is

waiting on anybody's gradient.

Owning the weights, setting the refusal policy, pricing access and keeping the logs are not interconnect problems. Those are choices, and they are the ones worth arguing with.

Inference runs on llama.cpp with GGUF weights, so CUDA, ROCm, Metal, Vulkan and plain CPU all work. Nothing in the delivery path needs one vendor's silicon.

The aim, which is a bet and not a shipped feature: an internet of intelligence that nobody owns. Unsolved parts are in OPEN_PROBLEMS.md.

HART = the bare engine (run from this repo, listens on

:6777). There is no published PyPI package yet, so install it from source as shown below. HART OS = the full AI-native OS. It boots on a laptop, server, or edge node, runs on phones, and reaches into embodied AI, and it ships the agentic Liquid Shell, Model Bus, model catalog, channel pairing, agent dashboard, and hive view. Nunba = the consumer companion app, one signed client across Windows / macOS / Linux.

AI-native means the OS adapts to the machine, not the other way around. On each device it probes what the hardware can actually do, serves LLM, vision, and speech to every app over the Model Bus (socket, D-Bus, or HTTP), and lets the on-device model compose the interface and learn each task once so it can replay it later. The runtime that drives a desktop is the same one that drives a robot, so a robot's AI access is just another Model Bus call. It is one Python codebase that runs in three shapes (flat laptop, regional LAN, or central cloud mesh), speaks the OpenAI protocol on :6777/v1/chat/completions, and federates with peers over PeerLink (direct peer-to-peer WebSocket, no broker). A boot-time guardrail hash, re-checked every 300 seconds, plus Ed25519 release signing, keep humans in control.

You would notice it last, the way you notice anything alive: it improves on its own. Each node learns from what it does and gets quietly better, locally, on your own hardware, with nothing leaving the device. Calling an operating system alive should make you reach for the off switch, so that came first: the self-improvement is a toggle, every node is killable on its own, and it runs only as long as you let it.

This README is written to be read by people and by agents alike. Every capability below names the file it lives in, so whether you are a developer or an AI agent exploring the repo, you can go from a feature straight to its source.

Status: public alpha. The runtime, the Model Bus and the channel adapters are in daily use; APIs still move. Issues and PRs are genuinely wanted — see Contributing.

Want something concrete to pick up? good first issue and help wanted are real gaps rather than manufactured onboarding tasks. Each says what is wrong, why it matters, and what would count as done. A few of them carry the measurement that found the bug, and one names a hypothesis we already ruled out so nobody spends a Saturday re-testing it.

If you would rather argue than patch, start at Open problems. Nine things we have not solved, each with the code that implements today's inadequate answer and what would count as progress: what convergence can mean with no global view, whether a system that rewrites itself can still be verified, and why a turn escalates itself to a better model automatically but can never decide on its own that a problem deserves an hour and three machines. Disagreeing with a framing there is worth more to us than a patch.

- What it is

- Why HART OS?

- Open problems — what we have not solved

- Start it

- How it compares

- Capabilities

- Hello, agent

- Architecture map

- API surface

- How auto-evolve works

- How hive connectivity works

- Topology

- Build / extend

- Economics (for node operators)

- Documentation index

- License