How to get a YouTube transcript into Claude

Claude can’t watch a YouTube video, and pasting the link doesn’t change that. It reads captions, so you hand it those: paste the transcript yourself, or add a connector that fetches it every time you drop in a link. For one video, pasting is fine. If you do this often, the connector takes two minutes to set up and the copying stops.

With only a link, Claude summarizes someone else’s summary

Give Claude only a link and it improvises. Asked in Claude Code to summarize a 35-minute Veritasium video, it first tried to download the captions with yt-dlp, a command-line tool it needs your permission to run. Then it fetched the YouTube page, which gave it the title and little else. So it searched the web, found a third-party summary site and summarized that.

It said so, but the section times it gave, 2:45, 9:40 and 18:00, came from the summary site. The video’s own chapters start at 3:09, 5:27 and 8:04. A summary built this way reads fine and sends you to the wrong minute. The Claude apps have the same gap: Claude reads text and images, and its uploads take neither video nor audio.

Add a transcript connector once, then just paste links

A connector gives Claude a tool it can call on its own. LurkAPI’s gives it youtube_transcript: you paste a link, Claude fetches the captions and answers from what was said.

claude.ai and Claude Desktop

1. Open Customize > Connectors, click Add custom connector and paste https://api.lurkapi.com/mcp?platforms=youtube. If it asks how Claude should identify itself, pick Register automatically.
2. Sign in on lurkapi.com with Google or an email code, then click Allow. New accounts get 100 free credits, and a lower balance tops back up to 25 every day.
3. In a chat, click + > Connectors and switch LurkAPI on.

From then on it works in the Claude mobile app too. Free plans allow one custom connector (Anthropic’s guide); on Team and Enterprise, an Owner adds it and members click Connect.

Claude Code

If Claude Code is signed in with the same Claude account, the connector is already there. If not, it’s one command. Keep the quotes: zsh, the default shell on a Mac, trips over a bare ?.

claude mcp add --transport http lurkapi "https://api.lurkapi.com/mcp?platforms=youtube"

Claude asks you to sign in the first time it uses it, or run /mcp to do it now. For scripts, skip the sign-in and pass a key with --header "x-api-key: $LURK_KEY".

A transcript costs 1 credit. It reads the captions, the creator’s own or YouTube’s automatic ones, in any language the video has them. Claude usually fetches the video’s details alongside (title, chapters, views) for 1 credit more, so a summary comes to 2 credits, and the 100 free credits cover about 50 summaries. A two-hour podcast arrives in three parts: you pay for the first, and Claude fetches the rest free for 15 minutes.

Then ask the way you’d ask a friend who watched it. The connector also reads comments and channels, so these all work:

• “Summarize this video with its chapter times.”
• “Where does she talk about pricing? I want to clip it.”
• “What do the top comments push back on?”
• “Summarize the last five uploads on this channel, one line each.” That’s one call for the list and a transcript per video, 6 credits.

Plain text for summaries, timestamps to find a moment

Timestamps take room. That 35-minute video is about 37,000 characters as plain text. As timed lines, with a start and an end on each of its 903 captions, it’s about 98,000. A summary doesn’t need the times; finding a moment does.

So the connector sends plain text by default, and the timed lines only when Claude asks for them. Ask, in the same chat, where the video talks about the turkey because you want to clip it, and Claude calls the tool again with timestamps on.

The turkey bit runs from 6:02 to 7:09, about 66 seconds. … The answer ends at 7:08.96 on “...not 100% longer.”

It cost 1 credit. If you paste a transcript instead, keep the timestamps for the same reason: without them Claude can quote a passage but can’t tell you where it is. Our free transcript tool has a Copy with timestamps button.

{
  "success": true,
  "credits_charged": 1,
  "_note": "Part of the result, to fit: transcript 1–413 of 903. For the rest (transcript, availableLanguages), call again with the same arguments and offset=413 (free for 15 minutes).",
  "transcript": [
    { "text": "- Really familiar example would", "startMs": 362520, "endMs": 364800, "startTimeText": "6:02" },
    { "text": "come from thinking about cooking.", "startMs": 364800, "endMs": 367666, "startTimeText": "6:04" },
    { "text": "(mellow music)", "startMs": 367666, "endMs": 367680, "startTimeText": "6:07" },
    { "text": "Maybe you want to make a big turkey for Thanksgiving.", "startMs": 367680, "endMs": 370320, "startTimeText": "6:07" }
  ]
}

Other ways to get the transcript

Local servers run on your computer, so claude.ai and the phone app can’t reach them. Run on a server, the same kind of fetch tends to meet YouTube’s bot check: why youtube-transcript-api gets blocked.

Building transcripts into your own app instead? The same transcript is one HTTP call: