Spoken MCP Server

Spoken — podcast transcripts as clean Markdown, built for AI agents

Spoken is a transcript API that turns any published podcast into clean Markdown with real speaker names — not "Speaker 1." One API call returns named, timestamped text, ready for LLMs, RAG pipelines, summarizers, and search.

It's a transcript retrieval API, not a speech-to-text service: it works on already-published podcasts, so you skip uploading audio, running diarization, and mapping anonymous speaker labels by hand. For published shows that's typically 5–10× cheaper than running the audio through a transcription service.

🎙️ Real speaker names, resolved automatically
📄 Clean Markdown with timestamps, tuned for LLM context windows and RAG chunking
🔎 Search by text query or paste a Spotify/YouTube URL
💳 Pay-per-use credits — no subscription, failed calls never charged, repeat fetches free
🤖 Agent-native — ships with an Agent Skill, agents.md, llms.txt, and an OpenAPI spec

Get a key at spoken.md — or try it free with the demo key pt_demo (search works fully; transcripts limited to the demo episode).

Quickstart

# 1. Find an episode (by text, or paste a Spotify/YouTube URL)
curl -s 'https://spoken.md/search?q=huberman+sleep' \
  -H 'x-api-key: pt_demo'

# 2. Fetch the transcript as Markdown
curl -s 'https://spoken.md/transcripts/1000651996090' \
  -H 'x-api-key: pt_demo'

The transcript comes back as Markdown with named speakers and timestamps:

**John Smith** (0:00)
Welcome to the show. Today we're talking about...

**Jane Doe** (0:15)
Thanks for having me.

Endpoints

Method & path	What it does	Credits
`GET /search?q={query or URL}`	Find episodes; returns `id`, `title`, `podcast`, `date`	0
`GET /transcripts/{id}`	Return the Markdown transcript	1 on first fetch, 0 on repeat
`GET /balance`	Current credit balance + usage history	0
`POST /buy`	New-key checkout (Stripe)	—
`POST /top-up?key={key}`	Returning-customer top-up (Stripe)	—

Auth is the x-api-key header. Responses include X-Credits-Remaining and X-Credits-Charged. See agents.md for the full error table and response shapes.

Examples

examples/podcast_summarizer.py — fetch a transcript and summarize it
examples/rag_pipeline.py — chunk a transcript for a vector store / RAG
examples/quickstart.sh — search → transcript in two curl calls

Use as an MCP server

This repo includes spoken-mcp, a Model Context Protocol server that exposes Spoken to MCP-compatible agents (Claude Desktop, Cursor, Cline, …). It provides three tools:

Tool	Description
`search_podcasts`	Find episodes by text or a pasted Spotify/YouTube URL
`get_transcript`	Fetch an episode's transcript as Markdown with real speaker names
`get_balance`	Check remaining credits

Add it to your MCP client config (e.g. Claude Desktop's claude_desktop_config.json):

{
  "mcpServers": {
    "spoken": {
      "command": "npx",
      "args": ["-y", "spoken-mcp"],
      "env": { "SPOKEN_API_KEY": "pt_your_key" }
    }
  }
}

SPOKEN_API_KEY defaults to pt_demo (search works fully; transcripts limited to the demo episode). Get a real key at spoken.md.

Run from source instead:

npm install && npm run build
SPOKEN_API_KEY=pt_your_key node dist/index.js

Use with AI agents

Spoken is designed to be called by agents. Point your agent at the Agent Skill (also served at https://spoken.md/.well-known/skills/spoken-md/SKILL.md), or hand it agents.md. The OpenAPI spec makes it easy to wrap as a tool for any function-calling or MCP-compatible client (Claude, GPT, Cursor).

Pricing

Pay-per-use credits, no subscription. New keys: 100 for $15, 500 for $50, 2,000 for $160. Machine-readable at spoken.md/pricing.md.

Documentation