Podcast MCP

Speaker-aware MCP server for searchable podcast transcripts.

Podcast MCP indexes podcast episodes from an RSS feed, transcribes them,
identifies speakers, stores semantic embeddings in Postgres/pgvector, and
exposes read-only search and retrieval tools through the Model Context
Protocol. Any podcast RSS feed can be
indexed; the code stays generic, while the AI Report
deployment is one concrete example.

How it works

Podcast MCP runs two separate processes. Ingestion is offline and heavy and may
use GPUs; the MCP server is online, lightweight, and read-only.

Process 1: Ingesting new episodes

flowchart TD
    A[RSS feed] -->|polled on a schedule| B[RSS sync]
    B -->|audio URL| C[Transcription<br/>Faster Whisper on RunPod GPU worker or local]
    C -->|transcript segments| D[Speaker diarization<br/>pyannote]
    D -->|anonymous SPEAKER_00 labels| E[Speaker name map<br/>LLM resolves labels to names + confidence]
    E -->|named transcript| F[Chunk + embed<br/>transcript chunks to OpenAI embeddings]
    F -->|episodes, speakers, segments, chunks| G[(Postgres + pgvector)]

Process 2: Calling the MCP server

flowchart TD
    A[Claude / Cursor / ChatGPT / any MCP client]
    A -->|MCP tool call<br/>stdio or authenticated HTTP| B[MCP server<br/>thin tool handlers]
    B -->|use cases: search, episodes, speakers| C[Application services]
    C -->|read-only interface| D[Repository<br/>Postgres + pgvector]
    D -->|search / transcript results| A

Try the hosted demo

This repository powers a production deployment for
AI Report. Over 170 hours of podcast episodes have
been indexed and new episodes are indexed automatically on release.

Public MCP endpoint:

https://ai-report.bramdehart.nl/mcp

Demo bearer token:

Bearer 9b55e1de7f3e0e8972377d3d9a77330929f6446bccf0927ccc648bb0d512018c

The demo endpoint is read-only and rate-limited.

Real-world examples

Point any MCP client at the demo endpoint and ask questions in natural language.
The tools let the assistant list episodes, search semantically, zoom in on a
timestamp, and filter by speaker.

Example 1: What is this podcast about?

What episodes did you find?
What are the latest topics that have been discussed?

The assistant lists recent episodes and their metadata to orient itself.

Example 2: Semantic recall

What was said about Satya Nadella.

A semantic search (search_podcast_transcripts) retrieves the transcript
chunks whose meaning is closest to the query, together with the episode,
timestamps, speaker, and similarity score.

Example 3: Speaker-aware search

What did Alexander say about Anthropic?

search_by_speaker narrows the search to chunks attributed to Alexander,
combining the speaker filter with semantic ranking.

Example 4: Timestamp context

Give transcript context around timestamp 12:34 of the episode about AI
agents.

get_transcript_around_timestamp returns the raw transcript segments around
that timestamp so the assistant can quote the exact wording.

Example 5: Speaker confidence matters

Who were the guests in last week's episode, and can we sure
know who said what?

The assistant lists the episode speakers (get_episode) and reports the
speaker_confidence for each attribution, using the guidance below.

MCP tools

• list_episodes — list indexed podcast episodes.
• get_episode — fetch episode metadata and speaker mappings.
• search_podcast_transcripts — semantic search over transcript chunks.
• get_transcript_around_timestamp — raw transcript context around a timestamp.
• search_by_speaker — search or list chunks by speaker.

Features

• Reads any podcast RSS feed and indexes only new episodes.
• Full offline ingestion pipeline: download, transcribe, diarize, name
  speakers, chunk, embed, store.
• GPU transcription on RunPod Serverless (or locally).
• Speaker-name resolution backed by an LLM with confidence scores and evidence.
• Semantic search with pgvector.
• Read-only, authenticated MCP server over stdio or HTTP.

Installation

Requires Python 3.11+.

python3.11 -m venv .venv
.venv/bin/pip install -e ".[server,ingestion]"
cp .env.example .env

Dependencies are split so a plain MCP server install stays small:

Install only what you need, e.g. pip install -e ".[worker]" on a GPU worker.

Quick start

Run the MCP server over stdio:

podcast-mcp

Query transcripts from the command line:

podcast-mcp-tools list-episodes
podcast-mcp-tools search "wat werd er gezegd over Anthropic?"

Ingest new episodes once, or on a schedule:

podcast-mcp-ingest        # one RSS sync
podcast-mcp-scheduler     # sync on SYNC_CRON

Start a local Postgres with pgvector:

docker compose up -d postgres

Speaker attribution

Speaker names are inferred from diarization and transcript context. Tool
consumers should account for speaker_confidence:

• >= 0.85: treat the speaker name as certain.
• 0.60-0.85: phrase attribution as likely or probable.
• < 0.60: mention that the speaker identity is uncertain.

Documentation

• Architecture
• Configuration
• Development
• MCP server and tools
• Deployment: Docker, RunPod, Hetzner
• AI Report deployment example
• Security

Tech stack

• Python
• MCP Python SDK
• Postgres + pgvector
• Docker Compose
• RunPod Serverless GPU workers
• Faster Whisper
• pyannote speaker diarization
• OpenAI embeddings (text-embedding-3-small) and speaker-name mapping
• Caddy for public HTTPS reverse proxy

Repository layout

src/podcast_mcp/
  server/           MCP server and thin tool handlers
  application/      use cases (search, episodes, speakers)
  domain/           entities (Episode, Speaker, TranscriptChunk, ...)
  infrastructure/   Postgres repository, embeddings, RunPod adapter
  ingest/           RSS sync, scheduler, transcript ingest, transcription

db/migrations/      Postgres schema
docs/               architecture, configuration, development, deployment
examples/ai-report/ concrete production deployment (AI Report demo)

Status

The code is intended to stay generic for any podcast RSS feed. The demo
deployment on AI Report is a concrete
implementation for one indexed podcast and lives in examples/ai-report/.