The Curator

An MCP server that generates AI-powered podcasts — on any topic, in minutes.

The Curator uses Google Gemini to write a natural, multi-turn podcast transcript and Vertex AI TTS to synthesize it into audio. It exposes two MCP tools that any MCP-compatible client (Claude Desktop, Zed, Cursor, etc.) can call directly.

---

MCP Tools

Tool	Description
create_podcast_transcript	Generates a podcast transcript between hosts Annabelle and Link on any topic
create_podcast_episode	Synthesizes a transcript into a .wav audio file and uploads it to GCS

Tech Stack

• Python 3.12
• FastMCP — MCP server framework
• Google Gemini (via Vertex AI) — transcript generation
• Vertex AI TTS — multi-speaker audio synthesis
• Google OAuth 2.0 — single-user authentication for MCP clients
• Google Cloud Storage — episode and OAuth state storage
• Cloud Run — serverless hosting
• Hatch — environment and task management
• Ruff, MyPy, Pytest — code quality

Project Structure

.
├── Dockerfile
├── pyproject.toml
├── src/
│   └── the_curator/
│       ├── main.py                  # FastMCP server, OAuth routes
│       ├── podcast_generation.py    # Transcript + audio synthesis
│       ├── auth/
│       │   └── provider.py          # Google OAuth provider
│       └── utils/
│           └── vertex_client.py     # Vertex AI client wrapper
├── terraform/                       # GCP infra (Cloud Run, GCS, Secret Manager)
└── tests/

Local Development

1. Install Hatch:

   pipx install hatch

2. Set required environment variables:

   export GOOGLE_CLIENT_ID=...
   export GOOGLE_CLIENT_SECRET=...
   export ALLOWED_EMAIL=you@example.com
   export SERVER_URL=http://localhost:8000

3. Run the server:

   hatch run start

4. Run checks:

   hatch run check

Run with Docker

docker build -t the-curator .
docker run --rm -p 8080:8080 \
  -e GOOGLE_CLIENT_ID=... \
  -e GOOGLE_CLIENT_SECRET=... \
  -e ALLOWED_EMAIL=you@example.com \
  -e SERVER_URL=http://localhost:8080 \
  the-curator

Deploy to GCP

Infrastructure is managed with Terraform. The Cloud Run service pulls credentials from Secret Manager.

cd terraform
terraform init
terraform apply \
  -var="podcast_service_image_uri=gcr.io/the-curator-496412/the-curator:latest" \
  -var="google_client_id=..." \
  -var="google_client_secret=..." \
  -var="allowed_email=you@example.com" \
  -var="server_url=https://your-cloud-run-url"

A GitHub Actions workflow handles image builds and deploys on push to main.

Endpoints

Path	Description
/health	Health check
/sse	MCP SSE transport (client entry)
/.well-known/oauth-authorization-server	OAuth discovery
/oauth2/callback	Google OAuth callback