Pasting a transcript into Claude works, but it's a chore: extract, copy, switch windows, paste, repeat. With the Scribefy MCP server, Claude fetches the transcript itself the moment you mention a video — and can search YouTube and inspect videos too. This guide shows how to set it up and what it unlocks.

What MCP gives you

The Model Context Protocol lets AI assistants call external tools. Scribefy ships an MCP server (scribefy-mcp) that adds four tools to Claude:

extract_transcript — pull a video's full timestamped transcript
search_videos — free-text YouTube search
get_video_metadata — title, channel, duration, caption tracks
get_related_videos — YouTube's "Up next" feed

Three of the four are free; only extraction spends credits. Together they turn Claude from "summarize this text I pasted" into a research assistant that can find and read videos on its own.

Setup (Claude Desktop)

You'll need a Scribefy API key from the dashboard (API + MCP plan). Then open Settings → Developer → Edit Config and add:

{
  "mcpServers": {
    "scribefy": {
      "command": "npx",
      "args": ["-y", "scribefy-mcp"],
      "env": { "SCRIBEFY_API_KEY": "sk_live_…" }
    }
  }
}

Restart Claude Desktop. The Scribefy tools appear in the tools panel. (Using a different assistant? See the Cursor and Windsurf setups, the ChatGPT guide, or the overview for all assistants.)

Things you can now ask Claude

Once the tools are live, you can just talk to Claude:

Summarize: "Summarize this video: youtu.be/… — give me the five key takeaways with timestamps." (See summarizing a video.)
Quote-find: "Pull every quote from this talk where she discusses pricing, with timestamps."
Repurpose: "Turn this tutorial into a step-by-step blog post outline."
Compare: "Search YouTube for three videos on vector databases, then summarize how their explanations differ."

For that last one, Claude chains the tools itself: search_videos to find candidates, get_video_metadata to check each, then extract_transcript on the ones worth reading.

A research pattern that scales

The most powerful workflow is search → filter → extract:

Ask Claude to search a topic. It calls search_videos and lists candidates — free.
Ask it to check which are worth reading. It calls get_video_metadata (duration, channel, captions) — free.
Tell it which to dig into. It calls extract_transcript only on those — the only step that costs credits.

Because metadata and search are free and extraction is cached, exploring a topic across many videos stays cheap — you only pay for the transcripts you actually use.

Why timestamps matter for AI

Scribefy returns transcripts segmented with timestamps. That lets Claude cite where something was said ("at 12:40 she argues…"), which makes summaries verifiable and quotes traceable back to the source video. It's the difference between "the video says X" and "here's exactly where."

Frequently asked questions

Does this work in ChatGPT or other assistants?

Any MCP-compatible host that supports stdio transport can run scribefy-mcp. Claude Desktop, Cursor, and Windsurf are the common ones; the server speaks standard MCP JSON-RPC.

Do the free tools really cost nothing?

Search, metadata, and related-videos are free. Only extract_transcript spends credits, and cached transcripts are free even then. See pricing.

What if a video has no captions?

Claude will report that extraction failed because no captions are available. See transcripts without captions for the details.

Want it working in your assistant? Set up the MCP server, then try asking Claude to summarize a video.