YouTube Analysis MCP Server

The Problem

YouTube holds an enormous amount of knowledge that AI agents cannot reach because they only read text. This MCP server lets any MCP-compatible agent analyze a video from a URL.

Before this, getting Claude to reason about a video meant finding a transcript, copy-pasting it, and hoping the timestamps survived. Tutorials, conference talks, interviews, and walkthroughs were all locked away from agents that could not watch.

The Interesting Technical Detail

Gemini can analyze YouTube URLs directly. No downloading, no chunking audio through Whisper, no transcription API. You pass the URL and Gemini handles the multimodal understanding.

Most other models are text-only and have no native way to consume video, so this server essentially proxies Gemini's capability and exposes it as MCP tools. That one detail is what makes the whole thing simple: the hard part is already solved by the model.

What It Does

Pass any YouTube URL and the agent can summarize the video, ask specific questions, pull the full transcript, extract frames at specific moments, and search within the video for a topic.

In practice that means Claude Code can answer "what does this talk say about distributed systems?" or "summarize this tutorial and give me the key steps" without you touching a transcript first. Summaries come in brief, medium, or detailed form with timestamps.

The Stack

The whole server is TypeScript, typed end to end, built on the Model Context Protocol with the Google Gemini API doing the video understanding.

It works with Claude Code, Claude Desktop, and anywhere else MCP is supported. You need a Gemini API key and nothing else.

The MCP Pattern

Instead of asking an agent to write glue code against an API, you hand it a tool that does exactly the thing. This is one of several MCP servers I have built on that idea.

The same pattern powers my Image Generation MCP server and AGINEAR for markdown-native project management. I wrote up the build in the blog post, and if you end up debugging your own server, the online MCP inspector I built connects to any remote server from the browser.