Rabbit R1 LiveKit Voice Agent

Talk to any AI model from your Rabbit R1. Use GPT-5.2, Claude, Gemini, Qwen3.5, or any other model as a real-time voice assistant on the R1 powered by LiveKit. By default, uses GPT-4.1-mini / Cartesia for voice.

Pair it with local-voice-ai to run everything locally on your own hardware w/ no cloud providers needed.

What This Is

SKILL.md is a structured prompt that any AI coding assistant can follow to build a full R1 voice agent project from scratch. It works with Claude Code, Cursor, Windsurf, Copilot, Aider, or any LLM-powered coding tool – just feed it the prompt and let it build.

The resulting project includes:

• A Python voice agent (scaffolded via lk app create) that runs STT -> LLM -> TTS
• A web server (web-server.py) that serves the R1 frontend and provides a LiveKit token endpoint
• An R1 “creation” (creation/index.html) – the 240x282px voice UI that runs on the device
• An install page (creation/install.html) with a QR code for sideloading onto the R1

How to Use

Install the Skill

npx skills add ShayneP/rabbit-r1-livekit-skill

This installs SKILL.md so your AI coding assistant can use it automatically. Works with Claude Code, Cursor, Windsurf, Copilot, Aider, and other tools that support the skills format.

Once installed, just tell your assistant what you want:

“Build me a voice assistant for my Rabbit R1 that answers trivia questions.”

In Claude Code, you can also invoke it directly:

/rabbit-r1-livekit

Manual Usage (Without Installing)

If you prefer not to install, copy the contents of SKILL.md into your assistant’s context (paste it into chat, add it as a reference file, etc.) and describe what you want.

What the Prompt Guides Your Assistant Through

1. Install the LiveKit CLI (lk) if needed
2. Scaffold the agent with lk app create --template agent-starter-python agent
3. Set up LiveKit Cloud credentials via lk cloud auth
4. Create the web server and R1 frontend files
5. Start all services and open an HTTPS tunnel
6. Provide an install link for the R1

Architecture

R1 WebView
  --> web-server.py (port 8081, HTTPS via tunnelmole)
    --> LiveKit Cloud (WebRTC SFU + Agent Dispatch)
      --> Python Agent (STT -> LLM -> TTS)

Flow:

1. User taps the R1 screen (required for mic access)
2. Frontend fetches a LiveKit token from /getToken
3. Token includes a RoomConfiguration that tells LiveKit which agent to dispatch
4. Frontend connects to a LiveKit room via WebRTC and publishes mic audio
5. LiveKit Cloud dispatches the named agent into the room
6. Agent processes audio through STT -> LLM -> TTS and publishes audio back
7. R1 plays the response through its speaker

Project Structure (After Running)

project-root/
  web-server.py              # Static file server + /getToken token API
  creation/
    index.html               # Voice agent UI for R1 (240x282px)
    install.html             # QR code page for R1 installation
  agent/                     # Scaffolded LiveKit agent
    src/agent.py             # Agent entry point
    pyproject.toml           # Dependencies (managed by uv)
    Dockerfile               # Production deployment
    tests/                   # Eval framework

Prerequisites

• Python 3.11+ with uv
• Node.js/npm (for tunnelmole)
• A LiveKit Cloud account (free tier available)
• A Rabbit R1 device (or a browser for testing)

Running Manually

If the project is already built and you want to run it yourself:

# 1. Install web server dependency
pip install livekit-api

# 2. Install agent dependencies
cd agent && uv sync && cd ..

# 3. Start the web server (terminal 1)
python web-server.py

# 4. Start HTTPS tunnel (terminal 2)
# Install first if needed: npm install -g tunnelmole
tmole 8081

# 5. Start the agent (terminal 3)
cd agent && uv run python src/agent.py dev

# 6. Install on R1
# Open https://YOUR_TMOLE_URL/creation/install.html in a browser
# Scan the QR code with your R1

Customization

R1 WebView Constraints

Key limitations to be aware of if modifying the frontend:

• HTTPS required – mic access (getUserMedia) is undefined over HTTP. Use tunnelmole or similar.
• No WebGL – the Flutter-based WebView only supports Canvas 2D.
• Screen size – 240x282 pixels exactly.
• Touch quirks – use document.body touchstart with e.preventDefault(). Inline onclick in dynamic HTML won’t fire.
• Mic publishing – use manual publishTrack() instead of setMicrophoneEnabled() for reliability.

R1 Native Bridges

The R1 WebView exposes these JavaScript globals:

• PluginMessageHandler.postMessage() – send messages to R1 system
• FlutterButtonHandler.postMessage() – button events
• CreationStorageHandler.postMessage() – persistent storage
• closeWebView – close the creation

Custom events on window: sideClick, longPressStart, longPressEnd, scrollUp, scrollDown