audio-mcp

Remote MCP server for audio processing, deployed to Aurora and reachable only over Tailscale.

What it does

• Transcribe audio from YouTube URLs, HTTP(S) audio URLs, inline base64 payloads, or previously uploaded files. Backends: Groq Whisper (cloud, default) or local faster-whisper on CPU.
• Generate audio from text, with three backends: piper (local, free, Polish default voice), Google Cloud TTS Standard, OpenAI gpt-4o-mini-tts. Polish text is normalised (URLs and long hashes removed, acronyms respelled phonetically).
• Job artefacts (transcription.json, transcription.txt, audio.mp3) are downloadable via HTTP using URLs returned in each tool response.

Connecting clients

All clients connect over Tailscale to https://audio-mcp.uaru-teeth.ts.net/mcp.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "audio-mcp": {
      "type": "http",
      "url": "https://audio-mcp.uaru-teeth.ts.net/mcp"
    }
  }
}

Claude Code

Edit ~/.claude/.mcp.json:

{
  "mcpServers": {
    "audio-mcp": {
      "type": "http",
      "url": "https://audio-mcp.uaru-teeth.ts.net/mcp",
      "timeout": 900000
    }
  }
}

The raised timeout covers long CPU transcriptions.

Development

uv sync
make dev           # run server locally
make test          # run tests with coverage
make lint          # ruff check

Requirements: Python 3.12, uv, ffmpeg.

Architecture

• FastAPI + FastMCP v3 — HTTP + MCP protocol
• Groq — cloud transcription (Whisper)
• faster-whisper — local CPU transcription
• piper — local TTS (Polish, free)
• Google Cloud TTS / OpenAI TTS — cloud TTS options
• SQLite (aiosqlite) — job and upload metadata
• Tailscale — secure network access