Hey 👋 — sharing something I've been working on lately, in case it's useful to anyone doing similar stuff.
Quick context: I use Claude a lot for digging through YouTube — conference talks, tutorials, the occasional 2-hour podcast. And the workflow was always the same annoying dance: open the video, click "Show transcript," copy the wall of text, paste it into Claude, ask my question. Every. Single. Time. And you lose the timestamps, so when Claude says "they mention X around the middle," you can't easily jump back and check.
At some point it bugged me enough that I built a little thing to fix it. It's an MCP server — basically a small program that hands Claude (or Cursor / Windsurf) a set of YouTube tools so it can just… do the fetching itself. I called it Scribefy. (Heads up: it's my own project, so take the plug with the appropriate grain of salt — but the how-it-works bits apply to any YouTube MCP server.)
Once it's plugged in, the assistant gets four tools:
search_videos
— search YouTubeget_video_metadata
— title, channel, length, whether it even has captionsget_related_videos
— the "up next" listextract_transcript
— the full transcript, So now I just paste a URL and say "summarize this with timestamps," or even "find me 3 videos on X and tell me where they disagree" — and it figures out the rest. No copy-paste.
It started as a "scratch my own itch" thing, but the use cases turned out way broader than I expected. A few that keep coming up:
Different people, same core trick: let the AI read the video so you don't have to sit through all of it.
It's an npx
package, same config in Claude Desktop, Cursor, and Windsurf:
{
"mcpServers": {
"scribefy": {
"command": "npx",
"args": ["-y", "scribefy-mcp"],
"env": { "SCRIBEFY_API_KEY": "sk_live_…" }
}
}
}
(Claude Desktop → Settings → Developer → Edit Config. Cursor → ~/.cursor/mcp.json
. Windsurf → ~/.codeium/windsurf/mcp_config.json
, then Refresh.) Restart and the tools show up. There's a key + a couple of free credits to mess around with.
It's not really about one transcript. It's that the assistant can research a whole topic on its own — search for candidates, glance at the metadata to throw out the 3-hour ones, then only pull transcripts for the handful worth reading. Search and metadata are free, so poking around a topic stays cheap and you only "spend" on the videos you actually use.
And the timestamps matter more than I expected — Claude can go "at 12:40 they argue…" and I can click straight there to check, instead of trusting a vibe-summary of a 40-minute video.
So — not magic, not for everyone. But if YouTube is part of how you learn or work, having the AI do the watching-and-pulling has been a genuinely nice upgrade.
It's pretty fresh — just got it listed in the official MCP registry and I'm slowly telling people. Honestly I'm still figuring out what folks actually want it to do, so if you try it (or you've built something similar), I'd love to hear what tools you'd want your agent to have for YouTube.
Anyway — that's the thing. Back to building. 🛠️