CHeema-Text-to-Voice-MCP-Server

AI-powered text-to-speech MCP server with instant voice cloning. Generate speech from Claude Desktop, Claude Code, or n8n using 5 built-in voices (English, German, French, Spanish) or clone any voice from a short audio sample. Runs fully local, no API keys, no cloud. Supports stdio, SSE, and HTTP transports.

Cheema Text-to-Voice is a free, open-source MCP server that gives any MCP-compatible AI assistant, including Claude Desktop, Claude Code, and n8n, a real voice. It runs the NeuTTS text-to-speech models locally on your own CPU or GPU, so there are no API keys to configure, no cloud service in the loop, and no per-character billing: text goes in, a WAV file comes out, entirely on your machine. Ask your assistant to speak and it synthesizes natural speech in five built-in voices across four languages, or clones a new voice from a short recording in seconds.

- MCP directory: listed onmcpservers.org
- Case study:how Cheema Text-to-Voice was built

Built byTayyab Ilyas, AI Agent Engineer in Barcelona.

# Ubuntu / Debian sudo apt install espeak-ng # macOS brew install espeak-ng # Windows choco install espeak-ng

git clone https://github.com/MuhammadTayyabIlyas/CHeema-Text-to-Voice-MCP-Server.git cd CHeema-Text-to-Voice-MCP-Server python -m venv venv source venv/bin/activate # Linux/macOS # venv\Scripts\activate # Windows pip install -e . pip install "mcp[cli]"

Pick your platform and follow the steps below.

- macOS:~/Library/Application Support/Claude/claude_desktop_config.json
- Windows:%APPDATA%\Claude\claude_desktop_config.json

{ "mcpServers": { "cheema-tts": { "command": "/full/path/to/CHeema-Text-to-Voice-MCP-Server/venv/bin/python", "args": ["/full/path/to/CHeema-Text-to-Voice-MCP-Server/mcp_server.py"], "env": {} } } }

Restart Claude Desktop. You'll see the TTS tools appear in the tools menu.

claude mcp add cheema-tts -- /full/path/to/venv/bin/python /full/path/to/mcp_server.py

Then just ask Claude to generate speech in any conversation.

cd CHeema-Text-to-Voice-MCP-Server source venv/bin/activate python mcp_server.py --transport sse --host 127.0.0.1 --port 8000

- Add anAI Agentnode with anMCP Client Tool
- Set connection type toSSE
- Enter the URL:http://127.0.0.1:8000/sse
- The agent can now call any TTS tool

Start the server with your preferred transport:

# SSE (for web platforms and remote clients) python mcp_server.py --transport sse --host 0.0.0.0 --port 8000 # Streamable HTTP python mcp_server.py --transport streamable-http --host 0.0.0.0 --port 8000

- SSE:http://<your-host>:8000/sse
- HTTP:http://<your-host>:8000/mcp

Usetts_list_speakersto see all voices including any custom ones you've added.

Clone any voice from a short audio sample:
- Record or find a WAV file: 3 to 15 seconds of clean speech
- Know the transcript: the exact words spoken in the recording
- Ask your AI assistant:

"Add a new speaker called 'alex' from /path/to/recording.wav, with the transcript 'This is what I said in the recording'"

tts_add_speaker(name="alex", wav_path="/path/to/recording.wav", ref_text="This is what I said in the recording")

Custom voices are saved permanently and available across restarts.

- Mono audio, 16-44 kHz sample rate
- 3-15 seconds of continuous, natural speech
- Minimal background noise

These appear automatically in Claude Desktop's prompt picker.

The default model (neutts-nano) works great on CPU. Larger models produce higher quality but need more resources.

Switch models using environment variables:

NEUTTS_BACKBONE="neuphonic/neutts-air" python mcp_server.py

All settings are optional. Defaults work out of the box.

NEUTTS_BACKBONE_DEVICE=cuda NEUTTS_CODEC_DEVICE=cuda python mcp_server.py

For production use, create a systemd service so it starts automatically:

# /etc/systemd/system/cheema-tts.service [Unit] Description=Cheema Text-to-Voice MCP Server After=network.target [Service] Type=simple WorkingDirectory=/path/to/CHeema-Text-to-Voice-MCP-Server ExecStart=/path/to/venv/bin/python mcp_server.py --transport sse --host 127.0.0.1 --port 8000 Restart=always RestartSec=5 [Install] WantedBy=multi-user.target

To expose it over HTTPS, put an Nginx or Caddy reverse proxy in front with these key settings for SSE:

- Disable proxy buffering (proxy_buffering off)
- Set a long read timeout (proxy_read_timeout 86400)
- AddX-Accel-Buffering: noheader

- Check espeak-ng:espeak-ng --version
- Check dependencies:pip list | grep -E "mcp|neutts|torch|soundfile"

- Check theoutput/directory for WAV files
- Verify speaker name withtts_list_speakers

- Normal: the first run downloads model weights from HuggingFace (~200-500MB). Cached after that.

- SetNEUTTS_BACKBONE_DEVICE=cudaandNEUTTS_CODEC_DEVICE=cuda
- The server loads the NeuTTS backbone model and audio codec on startup
- Speaker voice prints (.ptfiles) are loaded into memory
- When you request speech, text is phonemized and combined with the speaker's voice reference
- The model generates speech tokens, decoded into a 24kHz waveform
- Output is saved as a standard WAV file

CHeema-Text-to-Voice-MCP-Server/ ├── mcp_server.py # MCP server entry point ├── neutts/ # NeuTTS engine ├── samples/ # Built-in speaker voices (.wav, .pt, .txt) ├── speakers/ # Custom cloned voices (auto-created) ├── output/ # Generated audio files (auto-created) └── examples/ # Usage examples

- NeuTTSbyNeuphonic, the TTS engine
- MCP ProtocolbyAnthropic, the AI tool standard

MIT License. The underlying NeuTTS models have their own licenses. See theNeuTTS repositoryfor details.

Live crypto technical analysis MCP server — EMA, RSI, MACD, ATR, Bollinger Bands, TSS scoring, and Claude AI bull/bear debate via CoinGecko free API