One tool and no credential: a URL goes in, text comes back, and the model can quote it rather than paraphrase a description. Turning timestamps on is what makes the quote checkable — you get the moment as well as the words, with the sponsor read already stripped out.
A single-tool server: hand it a YouTube URL, a Shorts URL or a bare video ID and it returns the video's captions as text. A second tool appears if you supply a TwelveLabs key, for videos where the captions are not the point.
- A transcript from a full watch URL, a Shorts URL or a bare video ID
- The caption language chosen with lang, falling back to whatever the video does have when your choice is missing
- Timestamps in the output when you ask for them, so a passage can be cited at the second it appears
- Sponsorships and promotional segments filtered out by default from the chapter markers — turn strip_ads off to get the raw captions
- An optional analyze_video call that reasons over what is shown on screen rather than said, for demos, gameplay and b-roll with little speech
Node.js 18 or newer; no account and no API key for transcripts. The video-analysis tool is opt-in and needs TWELVELABS_API_KEY, and it takes a direct video file URL — a YouTube watch page serves HTML, not a video stream, so it will not work there. MIT licensed.
One command — npx -y @smithery/cli install @kimtaeyoon83/mcp-server-youtube-transcript --client claude