I Built an MCP Server That Turns YouTube Videos Into AI Skills

#mcp #claudecode #python #ai

Ever watched a 30-minute programming tutorial and thought "I wish I could just extract the useful parts"?

I built Glean - an MCP server that watches YouTube videos for you and converts them into structured skill files that Claude Code (and other AI assistants) can actually use.

The Problem

We've all been there:

You find a great tutorial on YouTube
You take notes, pause, rewind, pause again
A week later, you need that info and can't remember which video it was
Your AI assistant has no idea about that specific technique you learned

The Solution

Glean connects to Claude Code via MCP (Model Context Protocol) and gives it three new abilities:

Generate a skill from any YouTube URL
Search YouTube and rank videos by educational quality
Automatically find the best video on a topic and learn from it

The output is a clean markdown file with concepts, code examples, best practices, and common pitfalls - all extracted from the video.

How It Works

You: "Learn about FastAPI authentication from this video: youtube.com/watch?v=..."

Claude: *extracts transcript*
        *removes sponsor segments, "smash that like button", filler words*
        *structures into a skill file*
        *saves to ~/.claude/skills/*

You: "How do I implement JWT auth in FastAPI?"

Claude: *now actually knows the answer from the skill you just created*

Quick Demo

After installing, you can do things like:

Search YouTube for "kubernetes networking" videos by TechWorld with Nana

Glean will search, filter by that channel, and rank results by educational value (not just view count).

Or go fully automatic:

Generate a skill about React Server Components from the best tutorial you can find

The Technical Bits

Transcript fallback chain: YouTube captions → yt-dlp → Whisper (via Groq)
Smart cleaning: AI removes sponsors, self-promo, engagement bait, filler words
Ranking algorithm: Weighs relevance, educational signals, engagement ratio, channel authority
Channel filtering: Search within specific creators' content

Installation

# Clone it
git clone https://github.com/tauanbinato/glean.git
cd glean

# Install with uv
uv sync

# Add to Claude Code
claude mcp add glean \
  -e GLEAN_ANTHROPIC_API_KEY=your-key \
  -e GLEAN_GROQ_API_KEY=your-key \
  -- uv run --directory /path/to/glean glean-mcp

You'll need API keys from Anthropic and Groq (free tier works fine for Whisper).

What Gets Cleaned Out

The AI cleaner automatically strips:

"This video is sponsored by..."
"Don't forget to like and subscribe"
"Follow me on Twitter"
"Link in the description"
All the "um", "uh", "basically", "you know"

What stays: the actual technical content, code examples, explanations.

Why MCP?

Model Context Protocol is Anthropic's standard for giving AI assistants new capabilities. Instead of copy-pasting or hoping the model knows something, you can build tools that extend what it can do.

Glean is one example. The skills it generates become part of Claude's knowledge for your projects.