voicebox-voice-mcp-agent

Voice cloning uses the /voices/clone endpoint with a name, reference audio file, and target engine such as qwen-customvoice. Send a multipart form POST request and receive a voiceId usable in subsequent speech generation calls.

Voicebox runs an MCP HTTP server at http://127.0.0.1:17493/mcp exposing tools like synthesize_speech, clone_voice, and start_dictation. Add the server URL to your client's .mcp.json configuration to enable agent-driven voice workflows.

Voicebox includes seven engines: Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox Multilingual and Turbo, HumeAI TADA, and Kokoro. All inference runs locally by default, with multilingual synthesis across 23 languages.

Voicebox is local-first, so all TTS and Whisper STT inference runs on your hardware and audio never leaves your machine unless explicitly configured. Internet is only needed for initial model downloads from HuggingFace.

Startup failures usually mean port 17493 is already in use by another process. Check with lsof -i :17493, kill the conflicting process, or set VOICEBOX_PORT to a different port before restarting.

Set the HF_TOKEN environment variable for gated models, verify sufficient disk space, and check proxy settings. You can also manually download models with huggingface-cli into the local models directory.