An open-source AI-native operating system developed by Hevolve AI (Hertz AI) in India. Its full name is “Hevolve Hive Agentic Runtime OS,” and it is designed around running AI models on your own machine so that everything keeps working even with Wi-Fi turned off. What makes it distinctive is that inference is handled by the OS rather than by individual applications. Apps hold no API keys and simply ask the OS for inference, which removes the waste of every app loading the same model separately. Nodes coordinate peer-to-peer without a central broker, sharing spare compute with one another. It is licensed under Apache 2.0 and is currently in public alpha.
Key Features
- A fully local inference engine: It runs models entirely on your own hardware using llama.cpp and GGUF-format weights. Even with the network disconnected, what you type never leaves the machine
- The OS manages inference centrally: The OS exposes an OpenAI-compatible
/v1/chat/completionsendpoint that applications query. Per-app API key management becomes unnecessary, and duplicate model loading is avoided - Automatic placement based on your hardware: A VRAM manager classifies the installed GPU and dynamically distributes models across GPU and CPU. CUDA, ROCm, Metal, Vulkan and CPU inference are supported, so machines without a GPU can still run via CPU offloading
- Brokerless peer-to-peer coordination: Nodes discover each other through gossip and connect directly over WebSocket. Compute can be borrowed between machines without standing up a central server, with contribution scoring built in
- A vision agent that operates the screen by looking at it: It combines screenshot capture with pyautogui desktop control, letting it handle GUI applications the way a person would
- A resident coding copilot: Claude Code runs as the node’s own copilot, confined to a read-only filesystem sandbox
- Replaying work through recipes: A procedure taught once is saved as a “recipe” and simply replayed afterwards. The official documentation cites speedups of up to 90%
- System areas that AI cannot rewrite: Built on a read-only NixOS store, updates go through a signed pipeline of tests, audits and benchmarks before being distributed as a new generation. If health regresses, it rolls back automatically
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Open source | $0 | All features under the Apache 2.0 license. No subscription |
As long as you run it on your own machine, there is no additional cost for using models. Charges only arise if you also use external cloud APIs such as OpenAI or Groq, billed by those services. The project describes a future economic model that splits revenue from API usage, ads and agent work 90/9/1, but the official repository notes that payments do not yet settle end to end.
Pricing is current as of August 2026. Check the official repository for the latest information.
Pros & Cons
✅ Pros
- Input is never sent outside the machine, which suits confidential documents and internal company data
- No subscription costs; you can keep using it within the limits of the hardware you already own
- No need to hand out API keys per application ─ key management is consolidated in one place, the OS
- Multiple applications no longer load the same model redundantly, keeping memory use down
- Apache 2.0 open source, so you can read the implementation to verify it or modify it yourself
⚠️ Cons
- It is a public alpha, so the APIs are still changing and production use requires careful validation
- Inference speed and the size of usable models depend on your hardware; you should not expect parity with large cloud models
- Python 3.10–3.11 is required and 3.12 or later will not work. Setting it up assumes some command-line knowledge
- Screen control by the vision agent needs to be used with the risk of misoperation in mind
- The economic model (contribution scores and revenue sharing) is still partway between concept and implementation
Comparison with Similar Services
| Criteria | HART OS | Ollama | LM Studio | Jan |
|---|---|---|---|---|
| Positioning | AI runtime at the OS layer | Local LLM runner | Local LLM environment with a GUI | Local AI chat app |
| Primary usage | The OS serves inference to apps | Run models via CLI and API | Chat through a desktop app | Chat through a desktop app |
| Node coordination | Shares compute peer-to-peer | Standalone by default | Standalone by default | Standalone by default |
| Desktop control | Vision agent included | None | None | None |
| License | Apache 2.0 | MIT | Proprietary (free to use) | AGPL |
Ollama, LM Studio and Jan are all “tools for running models locally” that operate as a single application on top of an existing OS. HART OS sits one layer below that, embedding inference and an agent runtime into the OS itself, extending to multi-node coordination and signature verification of system updates. If you want to try things quickly, Ollama or LM Studio fit; if you want to build out the whole environment, HART OS is the option.
Who Is It For
- Anyone handling confidential data who does not want AI input leaving their machine
- People who would rather not stack up monthly AI subscriptions and prefer to keep everything on their own hardware
- Owners of several machines who want to use a spare GPU from another one
- Developers who want to read and understand the internals of a local AI environment, or modify it themselves
- Anyone who needs AI in environments where the network is unstable or cut off
Summary
HART OS is an open-source AI-native OS that moves inference from applications into the OS and pairs local-only operation with peer-to-peer coordination. Its design ─ no API keys handed to apps, automatic GPU/CPU placement, signed updates on a NixOS base ─ reads as an attempt to raise the local AI environment from a collection of tools into a foundation. That said, it is a public alpha today, and setting it up assumes a certain level of knowledge. Run it on your own machine first, check the inference speed and the day-to-day feel, and judge from there whether it is ready for everyday use.