HART OS Launches as Open-Source AI Operating System
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- HART OS launches as a public alpha — installable from source from the hertz-ai/HARTOS repo — with an OpenAI-compatible API on port :6777, exposing /v1/chat/completions so existing tools like LangChain, Aider, and Continue can drop in unchanged.
- Local inference runs through llama.cpp on GGUF weights, with vram_manager treating CPU-only, Metal, and broken nvidia-smi as ordinary cases; the project explicitly refuses the 'silicon lock' by supporting CUDA, ROCm, Metal, Vulkan, and plain CPU as first-class paths.
- Model federation uses PeerLink — direct peer-to-peer WebSocket with no broker — and the project deliberately rejects layer-level model sharding across consumer links, instead letting a peer with a bigger model serve a whole request.
- Safety guardrails include a boot-time hash re-checked every 300 seconds, Ed25519 release signing, an explicit kill switch per node, and a toggle that disables the self-improvement loop — described in the README as 'that came first: the self-improvement is a toggle, every node is killable on its own.'
- Install requirements demand Python 3.10 or 3.11 — a load-bearing constraint because faiss-cpu==1.7.4 publishes wheels only through cp311 — and requirements.txt pins 235 packages including torch, torchvision, transformers, onnxruntime, and scipy.
- The project ships 31 channel adapters plus integrations for MCP, Google A2A, Agent Protocol 2, AutoGen, and a CREATE/REUSE recipe system that replays cached task traces 90% faster without LLM calls on cached steps.
- An OPEN_PROBLEMS.md lists nine unsolved challenges — each naming the file that implements today's inadequate answer — including 'what convergence can mean with no global view' and whether a self-rewriting system can still be verified.
Why it matters: HART OS treats CUDA, ROCm, Metal, Vulkan, and CPU as first-class inference paths, so an Apple laptop, AMD card, or no-GPU box all run the same workload. The OpenAI-compatible endpoint on port 6777 lets existing tools drop in locally, while the project concedes pretraining frontier models still requires a datacenter — drawing an explicit line between the 'silicon lock' it breaks and the 'capital lock' it does not.
