MCP Tts Voicevox

What @kajidog/mcp-tts-voicevox does

@kajidog/mcp-tts-voicevox exposes VOICEVOX speech synthesis through MCP. An AI assistant can submit text, choose a speaker, adjust playback speed, and control whether playback starts immediately or waits for completion. Multiple dialogue segments can use different speakers in one request.

The server supports two playback paths. voicevox_speak plays generated audio on the server, while voicevox_speak_player creates an interactive MCP Apps player that sends playback to the client. The player is useful for remote connections and environments without an audio device on the server.

The player can manage multiple tracks, switch speakers, edit speech parameters, adjust accent phrases, and export numbered WAV files. It also includes a VOICEVOX user dictionary interface and restores player state across sessions.

How it works

The server communicates with a running VOICEVOX Engine instance. A standard local setup uses stdio and can be launched through npx or bunx; Claude Desktop starts the process from its MCP configuration. Docker Compose starts both the MCP server and VOICEVOX Engine, exposing the MCP endpoint over HTTP on port 3000.

The voicevox_speak tool accepts plain text or inline accent notation. Newline-separated segments can specify speaker IDs individually, allowing a single request to produce a multi-character exchange. When ffplay is installed and available on PATH, playback can use low-latency streaming; otherwise, the server falls back to temporary-file playback through the operating system's available audio player.

MCP Apps-compatible hosts can render the player UI in the conversation. The documented player clients are Claude Desktop and ChatGPT, with ChatGPT remote HTTP use requiring VOICEVOX_PLAYER_DOMAIN.

Setup and configuration

Install Node.js 24 or later, use Bun, or run the Docker setup. Start VOICEVOX Engine before launching the local MCP process. FFplay is optional, but installing FFmpeg enables streaming playback when ffplay is on the system path.

A Claude Desktop stdio configuration can launch the package with:

{
  "mcpServers": {
    "tts-mcp": {
      "command": "npx",
      "args": ["-y", "@kajidog/mcp-tts-voicevox"]
    }
  }
}

For Docker, run docker compose up -d, then connect Claude Desktop through mcp-remote at http://localhost:3000/mcp. The container has no audio device, so server-side voicevox_speak playback is disabled by default in that setup; use the client-side player instead.

Docker's published port has no authentication by default. Set MCP_API_KEY and send it with the X-API-Key header, or keep the port restricted to a trusted network. VOICEVOX_ALLOWED_OUTPUT_DIRS can limit directories used by file-writing tools.

Tools and capabilities

• voicevox_speak: Generate and play speech from text, with speaker, speed, queue, and completion controls.
• voicevox_speak_player: Create an interactive client-side audio player and return a view identifier.
• voicevox_resynthesize_player: Update every segment in an existing player session.
• voicevox_get_player_state: Read paginated state from a player session.
• voicevox_open_dictionary_ui: Open the VOICEVOX user dictionary manager.
• Edit speaker selection, speed, volume, intonation, pause lengths, silence, accent positions, and mora pitch in the player.
• Add, remove, reorder, preview, and export dialogue tracks as WAV files.

Limitations and notes

VOICEVOX Engine must be running for speech synthesis. Server-side playback also depends on an accessible audio device and playback utility; Docker does not provide an audio device. Hosts without MCP Apps support cannot use voicevox_speak_player, so voicevox_speak is the fallback.

The player can export WAV files even when opening the file explorer is unavailable. Selecting an output folder uses a native directory picker on Windows and macOS; unsupported environments use the default export directory. Remote or exposed Docker deployments should be protected because the default port configuration has no authentication.