YouTube Transcript MCP Server

Hosted — paste a URL, nothing to install

The same server runs at api.transcriptapi.io/mcp over Streamable HTTP. Put your key in the URL for clients that cannot send headers (claude.ai, ChatGPT), or send it as a Bearer header from clients that can. No OAuth dance, no local Node.

https://api.transcriptapi.io/mcp/ta_your_key

https://api.transcriptapi.io/mcp/ta_your_key

https://api.transcriptapi.io/mcp/ta_your_key

claude mcp add --transport http transcriptapi \
  https://api.transcriptapi.io/mcp \
  --header "Authorization: Bearer ta_your_key"

{
  "mcpServers": {
    "transcriptapi": {
      "url": "https://api.transcriptapi.io/mcp",
      "headers": { "Authorization": "Bearer ta_your_key" }
    }
  }
}

Local — run it with npx

{
  "mcpServers": {
    "transcriptapi": {
      "command": "npx",
      "args": ["-y", "transcriptapi-mcp"],
      "env": { "TRANSCRIPTAPI_KEY": "ta_your_key" }
    }
  }
}

claude mcp add transcriptapi \
  -e TRANSCRIPTAPI_KEY=ta_your_key \
  -- npx -y transcriptapi-mcp

{
  "mcpServers": {
    "transcriptapi": {
      "command": "npx",
      "args": ["-y", "transcriptapi-mcp"],
      "env": { "TRANSCRIPTAPI_KEY": "ta_your_key" }
    }
  }
}

{
  "servers": {
    "transcriptapi": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "transcriptapi-mcp"],
      "env": { "TRANSCRIPTAPI_KEY": "ta_your_key" }
    }
  }
}

{
  "mcpServers": {
    "transcriptapi": {
      "command": "npx",
      "args": ["-y", "transcriptapi-mcp"],
      "env": { "TRANSCRIPTAPI_KEY": "ta_your_key" }
    }
  }
}

TRANSCRIPTAPI_KEY=ta_your_key npx -y transcriptapi-mcp

The tools your assistant gets

Every tool accepts what a person would paste — a full URL with tracking parameters, a Shorts link, an @handle — and extracts the id itself, so the model never has to. All five are read-only and safe to auto-approve.

Translation on the transcript tool adds one credit per 40 caption segments, minimum two. Anything YouTube refuses — captions off, private, deleted, age-restricted — comes back as a plain-English error with retryable: false, and the credit is returned. Full pricing is on the pricing table; the REST endpoints behind each tool are in the API reference.

Things to ask once it is connected

• Summarize https://youtu.be/dQw4w9WgXcQ in five bullets, with the timestamp where each point is made.→ get_youtube_transcript → summary with [m:ss] citations
• What does this talk say about pricing? Quote it exactly.→ get_youtube_transcript → verbatim quotes, not paraphrase
• Find the three most-watched Veritasium videos about entropy and tell me which one to watch first.→ search_channel_videos → three transcripts → a recommendation
• List the last ten uploads of @lexfridman and pull the transcript of the newest one.→ get_channel_videos → get_youtube_transcript
• Turn this video into a German SRT file.→ get_youtube_transcript with translate_to: de, format: srt
• Go through this playlist and build me a study guide with one paragraph per video.→ get_playlist_videos → one transcript per video → guide

The transcript tool returns the text in 30-second blocks, each with one timestamp, so when the assistant cites [12:40] you can open the video at that second and check the claim against what was said.

Why an MCP server instead of letting the model fetch YouTube

Two reasons, and both are YouTube's. The official Data API does not return transcripts at all — it lists caption tracks but only lets the video's owner download them. And the unofficial route, reading the caption track the player uses, is blocked from every datacenter IP range an assistant or a CI runner is likely to have. A model that fetches YouTube itself works on your laptop and fails in production.

This server does none of that. It calls TranscriptAPI, which fetches through residential exits, caches every transcript so a second read is free, and turns YouTube's failures into structured errors the model can act on rather than retry. The server itself is a few hundred lines of TypeScript: parse the link, call the API, shape the response for a language model.

Frequently asked questions