VoiceStudio
Synthesizes speech from a 646-language catalogue, clones voices, and transcribes audio on hardware you control; transcription coverage depends on the ASR engine you enable. Needs a VoiceStudio backend running outside the bot behind an authenticated HTTPS proxy.
What it provides
Connect an HTTPS inference API. Model weights stay on CPU/GPU infrastructure and are never installed inside the 4 GB bot.
Connection requirements
Run VoiceStudio outside the bot on a host with its own RAM, disk, and optional GPU; it needs 8 GB RAM and 10 GB free disk at minimum, so it must never be installed inside the 4 GB bot. The MCP server is mounted on the running backend at /mcp on localhost:3900 — there is nothing extra to start once the backend is up. That transport ships UNAUTHENTICATED, so publish /mcp through a TLS reverse proxy that enforces the bearer token below; VoiceStudio itself only requires a share PIN or API key once you leave loopback, and the MCP endpoint carries no access control of its own. The backend also runs a DNS-rebinding guard that rejects unknown Host headers, so set OMNIVOICE_MCP_ALLOWED_HOSTS to the hosts your proxy presents or every call is refused. Renders come back as inline base64 WAV in the default resources mode, which is expensive in context, so keep each request short. The files mode that returns a path instead only helps when the backend and the caller share one mount at the same path, which a remote HTTPS connector does not — treat any server-side output_path as unreadable from here, and set OMNIVOICE_MCP_BASE_PATH only to bound file arguments, which are refused outright when it is unset. Budget for this connector’s fixed 120-second deadline covering the whole MCP exchange: a CPU-only host renders a paragraph in minutes and serializes generations, and raising the upstream OMNIVOICE_MCP_TIMEOUT_S or OMNIVOICE_GENERATE_TIMEOUT_S does not extend that deadline, so keep requests short enough to finish inside it and use a GPU host for anything longer. Voices, projects, and audio stay on that machine. The application is AGPL-3.0 and downloaded models keep their own upstream terms; only clone voices you have permission to reproduce.
- Protocol
- Streamable HTTP MCP
- Endpoint
https://voicestudio.example.com/mcp- Authentication
- bearer
mcp toolsmcp call list_voices {}Where to add it
Open Dashboard Tools, select a running OpenClaw or Hermes instance, then use the card’s Install or Connect action. External services may require an HTTPS endpoint and credentials; the dashboard shows those fields before anything is saved.