voicestudio-mcp

Official MCP server for Voice Studio. Voice Studio is a local desktop app for AI text-to-speech, voice cloning, voice design, and transcription. This package connects that app to your agents. Get the app at aiaudiogen.com.

An MCP (Model Context Protocol) server that lets agents drive Voice Studio, a local desktop app for text-to-speech, voice cloning, voice design, and transcription. Once installed, Claude Desktop, Claude Code, Cursor, and any other MCP-aware client gets 12 audio tools that talk to the Voice Studio backend running on your machine, plus a built-in usage guide so the agent knows how to use them well.

Requirements

• Voice Studio.app installed (download from https://aiaudiogen.com)
• Python 3.11+ (only needed if installing via pip - uvx ships its own)

The MCP server will auto-launch the Voice Studio app the first time you call a tool, then poll until the local backend is ready.

Install

Option A - uvx (recommended)

No global install. Each MCP client launches it on demand:

uvx voicestudio-mcp

Option B - pip

pip install voicestudio-mcp
voicestudio-mcp   # runs the stdio server (Ctrl+C to exit)

Option C - local checkout

cd integrations/mcp_server
pip install -e .

Wire it into your MCP client

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows) and add:

{
  "mcpServers": {
    "voicestudio": {
      "command": "uvx",
      "args": ["voicestudio-mcp"]
    }
  }
}

If you used pip install instead of uvx:

{
  "mcpServers": {
    "voicestudio": {
      "command": "voicestudio-mcp"
    }
  }
}

Restart Claude Desktop. The 12 vs_* tools will appear in the tool palette.

Claude Code

claude mcp add voicestudio uvx voicestudio-mcp

Or by editing .claude/mcp_config.json in your project:

{
  "mcpServers": {
    "voicestudio": {
      "command": "uvx",
      "args": ["voicestudio-mcp"]
    }
  }
}

Tools

Built-in guidance

Beyond the tools, the server exposes its own usage playbook so an agent knows how to drive Voice Studio well without any extra install:

• an MCP resource voicestudio://guide (Voice Studio usage guide), and
• an MCP prompt voicestudio_guide

covering recipes (voiceovers, cloning), timing (cold start, model swaps), and the free-tier generation limit. The same call-time essentials are baked into each tool's description, so a client that never reads resources still gets the important warnings.

Notes

• Audio is returned as file paths, not inline bytes. Generated WAVs can be 500 KB to several MB; we hand back the absolute path plus a one-line summary. Default output directory is ~/Documents/VoiceStudio/integrations-out/ (override with VS_OUTPUT_DIR).
• One ML pipeline at a time. Voice Studio loads one model into memory at any moment. Calling vs_generate_speech then vs_transcribe back-to-back will incur a model-swap latency while the first model unloads and the second loads.
• Auto-launch. If Voice Studio isn't running when you call a tool, the server spawns it and waits up to 45 seconds for the backend to answer /health (override with VS_LAUNCH_TIMEOUT).
• Backend port. Defaults to 127.0.0.1:58391. Override with VS_BACKEND_HOST and VS_BACKEND_PORT.

License

MIT