autonomous-os

What the robot runs

Autonomous OS presents itself as the open-source operating system for robots, described in the README as the Android for robots. The core idea is that a robot becomes autonomous once the OS is installed: everything it sees and hears is fed to an agentic reasoning engine running on the robot itself, and that engine decides what to do next. The reasoning engine is swappable. The README names six runtimes behind one AgentGateway interface with 76 methods: Hermes, OpenClaw, PicoClaw, Codex, Claude Code, and OpenCode. A user can switch the active engine live from the web UI, and the persona, memory, and connectors move with it. The robot acts through skills. Each skill is a folder containing one SKILL.md that the agent reads, and a skill acts by writing HW markers in its reply rather than touching hardware directly. Example skills include guarding the house, enrolling a face, following a person, reading facial mood, and setting a light, plus desk work such as Gmail, GitHub, and controlling a Mac. The robot also grows: a built in learning loop creates skills from experience, sharpens them with use, and searches its own past conversations. Every component is swappable, from the engine and model to the voice, skills, and board, and the robot declares what it has in a ROBOT.md file that the OS mounts exactly.

How the stack is layered

The system is a software stack in which each layer uses only the layer below it, so any layer can be replaced without touching the others. The top layer is apps: the Autonomous app adds a robot, sets up Wi-Fi, installs skills from a Skill Store, and switches brains, while the robot also serves its own setup and monitor UI from the web system on port 5000. Below that are skills, then the agentic runtime with its six engines behind the AgentGateway. The system services layer is a Go daemon called os-server on port 5000, with one package per box: intent answers fixed commands from a local table with no model, server strips HW markers out of a reply and posts them to the hardware abstraction layer before the words are spoken, agent switches engines, and bootstrap handles over the air updates as its own binary. Realtime voice runs Gemini Live, OpenAI Realtime, or Qwen inside the hardware layer beside the main path. The capabilities layer is the set of 13 names a robot may declare, such as audio, vision, sensing, motion, light, and display; ten of them mount HTTP routes on port 5001 with 111 endpoints and a live Swagger page. A safety gate below the engine clamps brightness, quiet hours, and explicit move speed from SAFETY.md with no model in the loop. Drivers are one folder per subsystem, and boards are one JSON entry per board, with Raspberry Pi 4, Pi 5, CM4, and OrangePi 4 Pro supported today.

Bringing your own robot

The project lets a user bring any robot they can describe in four markdown files. ROBOT.md is the body: the board and the hardware it has. SOUL.md is the self: who the robot is and how it talks. SAFETY.md is the bounds: how fast, how bright, how late. SKILL.md is the hands: one thing it can do. The README links a full guide for this path. The project ships concrete robots as references. The Autonomous Lamp is the flagship device that shows the whole OS and can see, hear, speak, and move. Reachy Mini is Hugging Face's desk robot, which runs the OS beside its own stack through a one command install script rather than a firmware flash. The Autonomous Intern is an always on desk agent with a mic, speaker, and LED ring. The contribution model favors skills as the easiest entry point: one markdown file, no Go and no hardware, and it lands on every robot that has the required parts. The README also lists paths for porting to a new robot, supporting new hardware through a driver class and a factory line, supporting a new board through a JSON entry, and adding a brain through an AgentGateway implementation. Licensing is split by folder: everything outside hal is Apache-2.0, and hal is GPL-3.0 by choice, so a driver committed there is GPL and a closed vendor SDK wraps out of process. The README notes the robot also carries other projects such as Pollen's Reachy Mini SDK, YOLOv8 for tracking, and Silero for voice, with full credit in a CREDITS file.

Editorial conclusion

The repository is published under the Apache 2.0 (with GPL-3.0 in hal/) license and its most recent commit was recorded on 2026-08-26. The source lives at https://github.com/autonomous-ai/autonomous-os.