YouTube MCP Server

Model Context Protocol server that pulls YouTube subtitles into any MCP-capable LLM client.

Free· Free and open source (MIT). You only pay for whatever LLM sits on the other end of the MCP connection (Claude Desktop, Claude API, or any other MCP-capable client).MCP Servers

Visit website →

Best for

Individual developers, researchers, and knowledge workers who use Claude Desktop or a similar MCP client and want a zero-friction way to summarise, translate, or extract quotes from YouTube videos.

Skip if

Teams building a production YouTube analytics pipeline, anyone who needs channel search or video-metadata queries, or environments where installing yt-dlp is not permitted.

YouTube MCP Server (published as @anaisbetts/mcp-youtube) is a tiny, focused Model Context Protocol server by Anaïs Betts that gives any MCP-capable LLM client the ability to pull the subtitle track from a YouTube video and hand it back as plain text. Under the hood it shells out to yt-dlp, so it inherits yt-dlp's ability to reach almost any public YouTube URL, including auto-generated captions when a creator has not uploaded their own. The server exposes a single tool that takes a video URL, downloads the closed-caption VTT/SRT, and returns the transcript to the model, which then does whatever the user asked for: a bullet summary, a chapter-by-chapter breakdown, a quote extraction, a translation, a tutorial-to-blog-post rewrite. Because the transport is standard MCP over stdio, it drops straight into Claude Desktop, Claude Code, Cline, Cursor, Windsurf, LibreChat, and any other client that speaks the protocol; setup is one config-file entry pointing at the npm package once yt-dlp itself is installed via Homebrew or WinGet. It is deliberately not a full YouTube Data API wrapper, there is no search, no channel enumeration, no metadata beyond what falls out of yt-dlp, and no video download of the actual video stream. The point is transcript-in, model-out, cheaply and locally, without touching Google's quota-limited APIs or handing an OAuth token to a third-party service. For anyone whose daily workflow includes 'I want to know what this video said without watching it,' it is the smallest possible thing that could work, which is exactly why it has become one of the most cloned reference MCP servers.

Editor's take

This is the MCP server I recommend first when someone asks 'give me a small, real example.' It does one thing, it does it well, and the whole source fits on a screen. If your workflow includes even one 'what did that video say' moment a week, install it tonight; if you need anything beyond captions, look for a heavier YouTube Data API server instead.

— The AI Tool Bible editorial team

Pros

✅ Truly minimal install: one npm package plus yt-dlp, one JSON entry in your MCP client, done in under two minutes
✅ No YouTube Data API key, no OAuth, no Google Cloud project, and no quota to burn through
✅ Works on any video yt-dlp can reach, including auto-generated captions and non-English tracks
✅ Runs entirely on your machine so transcripts never leave your box before the LLM sees them
✅ MIT-licensed reference implementation that is frequently forked as a starting point for other MCP servers
✅ Actively maintained by a well-known open-source developer (Anaïs Betts, author of mcp-installer)

Cons

⚠️ Only returns subtitles: no video download, no thumbnail, no channel or metadata queries, no search
⚠️ Requires yt-dlp installed and on PATH, which trips up users on locked-down corporate Windows machines
⚠️ Silently useless on videos that have no captions at all (rare, but happens on new uploads)
⚠️ yt-dlp itself is a moving target against YouTube's anti-bot measures, so occasional breakage until you update the binary
⚠️ No built-in rate limiting or caching, so summarising a 200-video playlist in a loop will get your IP throttled

Use cases

YouTube video summarisationTranscript extraction for note-takingTurning tutorial videos into written how-tosQuote and timestamp lookup across long talksNon-English caption translation via the LLMResearch on podcast and interview contentFeeding lecture transcripts into a study assistantReference implementation for building your own MCP server