local-voice

Claude Voice MCP

A Claude Code plugin that gives Claude a voice. After every response, Claude speaks a conversational summary aloud using Kokoro-82M, a high-quality neural text-to-speech engine that runs locally on your machine.

No API keys. No cloud services. Fully offline after setup.

Demo

Once installed, Claude automatically speaks after every response. You can also ask Claude to change voices, adjust speed, or list available voices naturally in conversation.

Setup

Prerequisites

• Claude Code CLI installed and working
• uv (Python package manager)

Platform-specific:

Platform	Extra requirement
Windows	Visual C++ 2015-2022 Redistributable (usually already installed)
macOS	brew install libomp (required by ONNX Runtime)
Linux	sudo apt install libportaudio2 (Debian/Ubuntu) or sudo dnf install portaudio (Fedora)

Install the Plugin

1. In Claude Code, run:

   /install-plugin https://github.com/nathanphelps/claude-voice-mcp
   

2. Run the install script to download TTS models (~337 MB) and verify dependencies:

   Windows (PowerShell):

   powershell -ExecutionPolicy ByPass -File install.ps1
   

   macOS / Linux:

   ./install.sh
   

   Important: The install script pre-downloads the Kokoro model files. Without this step, the first speak call will hang while downloading 337 MB in the background.

3. Restart Claude Code. The voice plugin activates automatically.

Available Voices

28 voices across two accents and two genders. Default is af_heart.

American Female

ID	Name
af_heart	Heart (default, highest quality)
af_alloy	Alloy
af_aoede	Aoede
af_bella	Bella
af_jessica	Jessica
af_kore	Kore
af_nicole	Nicole
af_nova	Nova
af_river	River
af_sarah	Sarah
af_sky	Sky

American Male

ID	Name
am_adam	Adam
am_echo	Echo
am_eric	Eric
am_fenrir	Fenrir
am_liam	Liam
am_michael	Michael
am_onyx	Onyx
am_puck	Puck
am_santa	Santa

British Female

ID	Name
bf_alice	Alice
bf_emma	Emma
bf_isabella	Isabella
bf_lily	Lily

British Male

ID	Name
bm_daniel	Daniel
bm_fable	Fable
bm_george	George
bm_lewis	Lewis

Usage

The plugin is always active — Claude speaks after every response automatically. No commands needed.

You can interact with the voice naturally:

• "Switch to a British male voice"
• "Use the voice am_michael"
• "List all available voices"
• "Speak a bit faster"
• "Read that last part again"

MCP Tools

The plugin exposes three tools that Claude calls automatically:

Tool	Description
speak	Synthesize and play text aloud. Accepts text, voice (optional), and speed (optional, default 1.0).
list_voices	List available voices. Filter by language: en (all English), a (American), b (British), all (everything).
set_default_voice	Change the session default voice by ID (e.g. am_michael).

Speed Control

Pass a speed multiplier to the speak tool:

• 0.8 — slower
• 1.0 — normal (default)
• 1.2 — faster

How It Works

The plugin runs as an MCP (Model Context Protocol) server. Claude Code launches it automatically when the plugin is active.

1. Claude generates a text response
2. An always-active skill directs Claude to call the speak tool
3. The speak tool synthesizes audio using Kokoro-82M (ONNX) and plays it through your speakers via sounddevice

All processing is local. Audio is generated by the Kokoro neural TTS model (~310 MB ONNX) running on CPU via ONNX Runtime. No GPU required.

Troubleshooting

First speak call hangs

Run the install script (install.ps1 or install.sh). The hang is caused by model files downloading on first use. The install script pre-downloads them.

No audio output

• Check that your system default audio output device is set correctly
• On Linux, ensure libportaudio2 is installed
• Test audio with: python -c "import sounddevice; print(sounddevice.query_devices())"

Model loading is slow on first call

The Kokoro model takes ~2 seconds to load into memory on the first speak call of each session. Subsequent calls are fast.

Credits