Clawatar 🎭

From Agent Intelligence to Interactive Intelligence. Give your AI agent a body.

A web-based 3D VRM avatar viewer with real-time animations, voice chat, and lip sync — built for OpenClaw.

Screenshots

Cute sakura UI, multiple background scenes, camera presets, emotion bar, and 162 animations. VRM model not included — bring your own!

Quick Start

git clone https://github.com/Dongping-Chen/Clawatar.git
cd Clawatar
npm install
npm run start

Open http://localhost:3000 and drop your .vrm model onto the page.

Features

🎭 Avatar & Animation

• 162 animations — wave, dance, think, laugh, shrug, and more (Mixamo VRMA)
• Facial expressions — happy, sad, angry, surprised, relaxed
• Idle behavior — avatar looks around, stretches, yawns when waiting
• Touch reactions — click the avatar for headpats, pokes, and silly reactions ✨

🌸 Beautiful UI

• Sakura/anime theme — cute pink glassmorphism panels
• Background scenes — Sakura Garden 🌸, Night Sky 🌙, Cozy Café ☕, Sunset 🌅
• Camera presets — Face, Portrait, Full Body, Cinematic with smooth transitions
• Quick emotion bar — 😊😢😠😮😌💃 one-tap expression + animation combos

🎤 Voice & Chat

• Audio-driven lip sync — mouth moves to actual speech audio
• Voice input — speak via your browser's microphone
• Voice output — ElevenLabs TTS (optional, requires API key)
• AI conversation — powered by OpenClaw (optional)
• Multi-device routing policy — action/expression is broadcast to all paired devices, while reply text/audio is routed only to the device that triggered the turn

🏠 3D Scene System (Blender Pipeline)

• 6 scenes — Cozy Bedroom 🛏️, Izakaya 🏮, Café ☕, Phone Booth 📞, Sunset Balcony 🌇, Swimming Pool 🏊
• Blender procedural pipeline — Python scripts generate geometry + materials + lights → Cycles render → GLB export
• Emissive-only materials — all scenes use Emission shaders for reliable rendering in Three.js
• Auto emissive lights — brightest emissive meshes automatically spawn PointLights
• Camera freedom — orbit ±135° inside scenes, configurable per-scene camera + exposure
• Activity modes — Study, Exercise, Chill with themed camera angles + animations
• Scene loader — loadRoomGLB() loads single GLB as entire environment with character lighting

📹 Virtual Meeting Avatar

• Join Google Meet / Zoom — avatar appears via OBS Virtual Camera
• Listen & respond — captures meeting audio via BlackHole → Whisper STT → OpenClaw AI → TTS
• Smart triggers — responds when called by name or asked a question
• Streaming pipeline (v3) — VAD + OpenClaw orchestrated model + streaming ElevenLabs TTS
• No direct LLM calls — all AI routes through OpenClaw Gateway (model selection, context, persona handled automatically)
• Rolling context — maintains 2-minute transcript window for coherent responses

🔌 Developer-Friendly

• Local WebSocket API — control everything programmatically on the same machine
• Drag & drop — load any VRM model
• Standalone mode — works without OpenClaw or ElevenLabs
• OpenClaw skill — install as an agent skill for AI-driven avatars

Bring Your Own Model

No VRM model is bundled. You can:

1. Drag & drop a .vrm file onto the viewer
2. Set a URL in clawatar.config.json → model.url
3. Enter a URL in the Model panel in the UI

Configuration

Edit clawatar.config.json:

{
  "model": { "url": "", "autoLoad": true },
  "voice": {
    "elevenlabsVoiceId": "your-voice-id",
    "elevenlabsModel": "eleven_turbo_v2_5"
  },
  "server": { "vitePort": 3000, "wsPort": 8765, "audioPort": 8866 },
  "openclaw": { "gatewayPort": 18789, "sessionId": "vrm-chat" }
}

WebSocket Protocol

{"type": "play_action", "action_id": "161_Waving"}
{"type": "set_expression", "name": "happy", "weight": 0.8}
{"type": "speak", "text": "Hello!", "action_id": "161_Waving", "expression": "happy"}
{"type": "reset"}

Multi-device message split

{"type":"sync","category":"action","payload":{"actionId":"161_Waving","expression":"happy","expressionWeight":0.8}}
{"type":"speak_audio","text":"Hello!","audio_url":"https://...","audio_device":"<source_device>","target_device":"<source_device>","reply_device":"<source_device>"}

• sync/action is broadcast to keep avatar motion synchronized across all devices.
• speak_audio / audio_start / audio_chunk / audio_end are reply-routed to the focused source device.

Architecture

Browser (localhost:3000)
├── Three.js + @pixiv/three-vrm
├── VRMA animation playback
├── Audio-driven lip sync
└── Chat UI + Emotion Bar
    │
    │ Local WebSocket (loopback ws://127.0.0.1:8765)
    ▼
WS Server (server/ws-server.ts)
├── Command relay & routing
├── ElevenLabs TTS
├── OpenClaw Gateway bridge (all AI routing)
└── Meeting speech → Gateway API → orchestrated model
    │
    ▼
OpenClaw Gateway (localhost:18789)
├── Model orchestration (Opus/Sonnet/Codex)
├── Session & context management
└── Persona & memory

Apple App Transport Policy (Relay-only)

• iPhone, iPad, and macOS clients use relay transport only (/ws/client).
• Simulator builds follow the same relay-only policy.
• Direct app WebSocket transport (ws://127.0.0.1:8765) is removed from Apple clients.
• ws-server.ts binds loopback (127.0.0.1) by default and also rejects non-loopback WS clients as defense in depth. Set CLAWATAR_ALLOW_REMOTE_WS_CLIENTS=1 only for explicit LAN debugging.
• Pairing tokens are long-lived; add new devices with /pair/add-device instead of creating a new session.
• Model orchestration path is unchanged: relay bridge -> ws-server.ts -> OpenClaw gateway (:18789).

OpenClaw Skill

Clawatar includes an OpenClaw skill at skill/SKILL.md. Install it to let your AI agent control the avatar with animations, expressions, and voice.

Scripts

Building Scenes (Blender Pipeline)

Each scene is a Blender Python script that generates procedural geometry → exports GLB.

# Build a scene
/Applications/Blender.app/Contents/MacOS/Blender --background --python blender/build_izakaya_v4.py

# Copy to public
cp /tmp/izakaya.glb public/scenes/izakaya.glb

# Load in viewer
open http://localhost:3000?room=izakaya

Scene scripts (in blender/)

Key rules for scene scripts

• All emission strengths ≥ 3.0 — sub-1.0 gets baked dark by glTF exporter
• Use Emission shader only (not Principled BSDF) for reliable Three.js rendering
• Cycles renderer — 64 samples + denoiser
• Center stage clear — character stands at origin (0,0,0)
• Background elements at Blender -Y — they end up behind the character in Three.js
• GLB under 8 MB — optimize mesh complexity
• See SCENES.md for detailed scene configs and review scores

Virtual Meeting Setup

1. Install OBS Studio and BlackHole 2ch
2. Create a Multi-Output Device (Audio MIDI Setup) → your speakers + BlackHole 2ch
3. Set system output to the Multi-Output Device
4. OBS: Add Browser Source → http://localhost:3000?embed → Start Virtual Camera
5. Start the avatar: npm run start
6. Start the meeting bridge: npm run meeting:v3
7. In Google Meet: select OBS Virtual Camera (video) and BlackHole 2ch (mic)

See virtual-meeting/README.md for detailed architecture docs.

Credits

• Animations: Mixamo — non-commercial use, credit required
• VRM rendering: @pixiv/three-vrm
• Inspired by: moeru-ai/airi

License

MIT — see LICENSE