VoiceStudioMCP Server

VoiceStudio ships an MCP server inside the desktop app, which makes it an unusual entry in this catalogue: it is a speech backend an agent can call, but nothing leaves the machine and there is no key to provision. If VoiceStudio is running, the server is already mounted at /mcp on its local backend (http://localhost:3900/mcp by default) — there is no separate process to start.

The tools

• generate_speech — text to a WAV file, using either the voice bound to the calling agent or a named profile.
• clone_voice — create a new voice profile from reference audio.
• transcribe — audio to text, across the 646 languages the app supports.
• list_voices, list_personalities, list_languages — enumerate what is available locally.
• check_health — backend status and which GPU device is active, which is the fast way to tell whether a slow generation is falling back to CPU.

Per-agent voice binding is the design detail worth knowing. Clients pass an X-VoiceStudio-Client-Id header (or OMNIVOICE_CLIENT_ID through the stdio shim), and the app binds a voice to that identifier — so a coding agent, a writing agent and a dictation client can each keep their own voice without passing a profile name on every call.

Modern clients connect over HTTP directly to the mounted endpoint. Clients that only speak stdio use the bundled shim, python -m backend.mcp_shim, which proxies to the same HTTP endpoint and forwards the client id as a header.

Because it is local, the practical constraints are your hardware rather than a quota: 8 GB RAM minimum with 16 GB recommended, optional GPU acceleration via CUDA, Apple MPS/MLX or ROCm, and macOS 13.3+ on Apple Silicon, Windows 10/11 x64 or Linux x86_64. The app is AGPL-3.0; individual model weights keep their upstream licences.