← Back to demo

Set up StudioEngineer MCP

Runs locally on your computer and talks to your coding agent over stdio. Pick your system, then your agent.

1 · Install

  1. 1.Install FFmpeg, Git and uv

    uv also installs the right Python (3.11+) for you. Open a new terminal afterwards so the commands are on your PATH.

    # Homebrew (https://brew.sh)
    brew install ffmpeg git
    curl -LsSf https://astral.sh/uv/install.sh | sh
  2. 2.Check the tools

    ffmpeg -version
    uv --version
  3. 3.Get the project and install dependencies

    Clone (or download) the project, then install its locked runtime dependencies in the mcp-server folder.

    git clone https://github.com/daniyallko/code-to-ai-podcast ~/studio-engineer-mcp
    cd ~/studio-engineer-mcp/mcp-server
    uv sync --locked
  4. 4.Copy the absolute folder path

    You'll paste this into the agent config below.

    pwd

2 · Connect your agent

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

{
  "mcpServers": {
    "studio-engineer": {
      "command": "uv",
      "args": [
        "--directory",
        "/Users/you/studio-engineer-mcp/mcp-server",
        "run",
        "--no-sync",
        "python",
        "-m",
        "src.server",
        "--transport",
        "stdio"
      ],
      "env": {
        "TTS_PROVIDER": "edge"
      }
    }
  }
}
  1. Open Claude Desktop → Settings → Developer → Edit Config.
  2. Paste the block (merge with any existing "mcpServers").
  3. Fully quit Claude (tray/menu bar → Quit) and reopen it.

Check: Click the tools/plug icon in the chat box — studio-engineer should list 4 tools.

3 · Use it

Ask your agent in plain language, or use the built-in prompts pr-roast-podcast and incident-postmortem-war-room.

Use the studio-engineer tools to make a short 2-host podcast roasting this PR:
"Removed all retries from the payment client to cut latency."
Alternate Alex and Sam, add one interruption (negative overlap_ms),
a record_scratch SFX, then render_master_mix at -16 LUFS.

The finished episode is saved to mcp-server/renders/<session_id>/master.mp3.

Optional

Development validation

Optional after connection: install the test dependency and run the deterministic offline suite.

uv sync --locked --extra dev && uv run --no-sync pytest -q

If Kokoro is installed, include both extras so validation preserves it:

uv sync --locked --extra dev --extra kokoro && uv run --no-sync pytest -q

Test without an agent (MCP Inspector, needs Node.js)

npx @modelcontextprotocol/inspector uv --directory "/Users/you/studio-engineer-mcp/mcp-server" run --no-sync python -m src.server

Real voices

The configs use edge: free neural voices that send generated dialogue text to Microsoft (no API key). Override the cast with TTS_VOICE_ALEX and TTS_VOICE_SAM. For ElevenLabs set TTS_PROVIDER=elevenlabs and ELEVENLABS_API_KEY in the env block; its service receives generated dialogue text. For local Kokoro speech install its locked extra, download kokoro-v1.0.onnx and voices-v1.0.bin into mcp-server, and set TTS_PROVIDER=kokoro with its model paths.

uv sync --locked --extra kokoro

Troubleshooting

  • "uv" not found — GUI apps may not see your shell PATH. Replace "uv" with its full path (which uv).
  • Render fails — run ffmpeg -version; reinstall FFmpeg and restart the agent.
  • Server not listed — check the JSON is valid, the path is absolute, then fully restart the agent.