Spaces:
Running
title: Autonomous OS
emoji: 🤖
colorFrom: blue
colorTo: purple
sdk: static
pinned: true
thumbnail: https://cdn.autonomous.ai/production/ecm/260701/thinking2.webp
short_description: Open-source OS for physical AI agents
Open-source operating system for physical AI agents.
Autonomous OS turns edge hardware into a complete AI agent — one that sees, hears, speaks, moves, senses, remembers, and updates itself over the air. The OS handles everything below the brain: hardware drivers, WiFi provisioning, a web dashboard, OTA updates, MQTT bridge, multi-channel messaging (Telegram, Slack, Discord, WhatsApp), and a 2-second sensing loop that runs vision, motion, sound, and light detection in parallel.
The brain is a swappable agentic runtime — five built-in options (OpenClaw, Hermes, PicoClaw, Codex, Claude Code) covering Claude, GPT, Gemini, and Qwen backends. Each runtime reads plain-text SKILL.md files and calls device APIs directly — no MCP wrapper needed. Realtime voice runs natively via Gemini Live, OpenAI Realtime, or Qwen Omni for sub-second speech-to-speech; complex requests delegate back to the main agent.
Hardware is a plugin: missing peripherals are skipped at init, so the same OS image runs a 5-DOF desk lamp and a quadruped robot alike. Skills address capabilities, not hardware — a "come here" skill works on any device that can move.
Autonomous Lamp — 5-DOF expressive desk robot. Audio, vision, motion, light, display, and sensing. The first reference device.
Autonomous Intern — always-on desk agent. Audio, sensing, and light. Pure voice + ambient intelligence.
Unitree Go2-W — a different manufacturer's mobile robot running the identical OS. The Android playbook: same skills, different hardware.
25 built-in skills — guard, mood, habit, wellbeing, music, camera, LED, servo, display, emotion, sensing, face enrollment, voice, computer-use, connectors, and more. Each skill is a plain-text SKILL.md the runtime reads and executes via HTTP.
Realtime voice (S2S) — native speech-to-speech via Gemini Live, OpenAI Realtime, or Qwen Omni. Sub-second casual chat on the realtime model; complex tasks auto-delegate to the main agent. Local VAD + intent matching bypass the LLM entirely for known commands (~50 ms).
5 agent runtimes — OpenClaw (Claude, full-featured), Hermes (Claude, 50% less memory), PicoClaw (Haiku, runs on Pi Zero), Codex (GPT-4), Claude Code (CLI-native). Swap via one config field.
Full system, not just a runtime — WiFi setup wizard, React web dashboard, OTA bootstrap worker, MQTT bridge, multi-channel messaging, JSONL session memory with auto-summarization, and a 2-second sensing loop (motion, face, light, sound) that routes events to the agent.