@three-ws/audio-mcp

Give a 3D AI agent a voice and a face — text-to-speech, speech-to-text, audio-to-face lipsync, and the motion-capture clip library, from any MCP client.

A Model Context Protocol server that exposes the three.ws voice and animation pipeline over stdio. Make an avatar speak, let users talk back, turn that speech into a per-frame lipsync track, and browse the recorded motion-capture library — everything you need to drive a 3D agent's mouth and movement.

The speech, recognition, and lipsync lanes run on free NVIDIA NIM models (Magpie TTS, Riva ASR, Audio2Face-3D), with OpenAI as the paid TTS backstop — the same provider policy as the three.ws platform. No key required to start; point THREE_WS_BASE at a deployment and go.

#Install

npm install @three-ws/audio-mcp

Or run with npx (no install):

npx @three-ws/audio-mcp

#Quick start

Claude Code, one line:

claude mcp add audio -- npx -y @three-ws/audio-mcp

Claude Desktop / Cursor (claude_desktop_config.json or mcp.json):

{
	"mcpServers": {
		"audio": {
			"command": "npx",
			"args": ["-y", "@three-ws/audio-mcp"]
		}
	}
}

Inspect the surface with the MCP Inspector:

npx -y @modelcontextprotocol/inspector npx @three-ws/audio-mcp

#Tools

All five tools are read-only: they transform input (synthesize/recognize/animate) or read live library data, never mutating platform state. None are idempotent — the audio bytes, transcript, and catalog move between calls.

#Input parameters

text_to_speech — text (required, 1–4096 chars), voice (default nova), format (mp3|opus|aac|flac|wav|pcm, default mp3), model (OpenAI backstop only), language (BCP-47, default en-US), speed (0.5–2.0, OpenAI lane only).

speech_to_text — audio (required, base64 or data: URL), format (wav|pcm|flac|ogg, default wav), language (BCP-47, default en-US), sampleRate (Hz, PCM only), words (bool), model (advanced).

audio_to_face — text or audio (one required), format (wav|pcm, audio path), sampleRate (PCM only), voice / language (text path).

motion_capture_clips — kind (face|pose|hand|vmc), include_public (bool, authenticated only), limit (1–100, default 50), cursor.

motion_capture_clip — id (required, clip UUID).

#Examples

// text_to_speech
> { "text": "Hello, I'm your three.ws agent.", "voice": "nova" }
{
  "ok": true,
  "voice": "Magpie-Multilingual.EN-US.Aria",
  "model": "magpie-tts-multilingual",
  "format": "wav",
  "mime": "audio/wav",
  "sizeBytes": 274476,
  "audio": "data:audio/wav;base64,UklGR…"
}

// audio_to_face (text path — returns audio + the lipsync track together)
> { "text": "Watch my lips move." }
{
  "ok": true,
  "audio": { "base64": "UklGR…", "contentType": "audio/wav", "format": "wav" },
  "animation": {
    "fps": 30,
    "blendShapeNames": ["jawOpen", "mouthClose", "…"],
    "durationSec": 1.9,
    "frames": [ { "t": 0.0, "w": [0.0, 0.1, /* … */] } ]
  }
}

A face frame is { t, w }: t = seconds from clip start, w = weights in the order of blendShapeNames (ARKit-52 naming). Play the audio and sample the track by the audio element's currentTime.

#Notes on payloads

Audio crosses the wire as base64 (TTS returns it, ASR/A2F accept it). Both the synthesis input (text) and the recognition input (audio, ~8 MB cap) are bounded by the platform — keep clips to a spoken utterance rather than long recordings. For lipsync on audio you already have, pass wav/pcm directly; browser-recorded WebM/Opus must be decoded to WAV or PCM first (Riva and Audio2Face do not accept the WebM container).

#Requirements

• Node.js >= 20.
• Network access to https://three.ws (or your own THREE_WS_BASE).

#Environment variables

#Links

• Homepage: https://three.ws
• Changelog: https://three.ws/changelog
• Issues: https://github.com/nirholas/three.ws/issues
• License: Apache-2.0, see LICENSE

_{Part of the three.ws SDK suite — 3D AI agents, on-chain identity, and agent payments.
Website · Changelog · GitHub}