SuperClonica text to speech

SuperClonica MCP server

Text to speech for Claude and other MCP clients, running on your own computer: offline after the first run, no limits, 31 languages. It is the local sibling of the cloud server at https://superclonica.com/mcp, with the same tool names. Voices come from the SuperClonica catalog.

Works on Windows, macOS and Linux with Python 3.10 or newer. Chinese is not supported.

Set it up

You need uv, which gives the uvx command. The first start installs the engine, which takes a minute or two. The first speak downloads the voice model (about 400 MB) once.

Claude Desktop. Add this to claude_desktop_config.json and restart the app:

{  "mcpServers": {    "superclonica": {      "command": "uvx",      "args": ["superclonica-mcp"],      "env": { "SUPERCLONICA_API_KEY": "sc_..." }    }  }}

If the app cannot find uvx, put the full path to it in command. On Windows where uvx prints it, on macOS and Linux which uvx.

Claude Code:

claude mcp add superclonica -e SUPERCLONICA_API_KEY=sc_... -- uvx superclonica-mcp

Cursor (~/.cursor/mcp.json) takes the same JSON as Claude Desktop. VS Code (.vscode/mcp.json) takes the same entry under "servers", with "type": "stdio" added.

Without uvx, these start the same server: pipx run superclonica-mcp, or python -m superclonica_mcp after pip install superclonica-mcp.

The API key is optional. Stock and free catalog voices need none. Installing a bought or your own voice does; creating a key is free.

Tools

voice is an id from list_voices, a stock name (F1–F5, M1–M5) or an exact name. When a name fits several voices you get the candidates to choose from, never a guess.

To use your own voice file, put its JSON in the voices folder. It goes by its file name.

Buying a voice

1. Ask for a voice, for example "find a calm narrator". The assistant calls list_voices.
2. If it must be bought, the assistant calls buy_voice and gives you a link that opens the purchase window.
3. Confirm the purchase in your browser. Only your click can buy.
4. Say "install it". After that the voice works offline.

Settings

All are optional environment variables.

Faster on an NVIDIA GPU

uvx superclonica-mcp installs the CPU engine, and the CPU and GPU engines cannot share one environment. For the GPU engine (Windows or Linux x86-64, Python 3.11+), make an environment of its own:

uv venv superclonica-gpu --python 3.12uv pip install --python superclonica-gpu superclonica-gpu-cudauv pip install --python superclonica-gpu --no-deps superclonica-mcpuv pip install --python superclonica-gpu "mcp>=2.3,<3" "truststore>=0.10"

Then use the Python of that environment as the command, with -m superclonica_mcp as the arguments and SUPERCLONICA_DEVICE=gpu in env.

When something goes wrong

• "The voice model is still loading": the first run is downloading the model, and that goes on in the background. Ask again in a minute.
• The client shows no tools: look at the server's log (Claude Desktop: Settings → Developer). The server writes its log only to stderr.
• A long text times out: on a CPU, voice long texts in parts.

License

Apache-2.0. The voice model is Supertonic 3 under its own OpenRAIL-M license, downloaded from its original place by the engine.