A YouTube transcript MCP server for Claude, Cursor and Codex
Let your AI search the library and cite timestamps.
Capture YouTube videos with transcripts, frames, and comments. Keep everything on your disk as Markdown + JSON, with a local search index.
Start with a saved source
Save a YouTube video, then run uoink connect claude-code, uoink connect cursor or uoink connect codex and add the printed entry to your client. Follow the installation steps.
Windows 10/11, plus a Mac preview for Apple Silicon (macOS 15 or later). ChatGPT and claude.ai: copy and paste only; Uoink does not connect to the ChatGPT app over MCP.
Ask your connected AI
Search my saved YouTube videos for YOUR_TOPIC. Get an evidence card for the video I choose and cite its transcript timestamps.
Search, then get the evidence
- 01
search_clipsReplace YOUR_TOPIC with what you want to find.
{ "name": "search_clips", "arguments": { "query": "YOUR_TOPIC" } } - 02
get_evidence_cardChoose a source from the search results. Replace VIDEO_ID_FROM_SEARCH with its video_id.
{ "name": "get_evidence_card", "arguments": { "video_id": "VIDEO_ID_FROM_SEARCH" } }
Example calls use placeholders. They are not search results.
Check before quoting
search no longer requires every word. search_library, search_uoinks, and search_clips take optional match ("forgiving" by default, "strict" to require every word) and mode ("auto" by default, "keyword", "semantic", "hybrid"); find_mentions takes match. Auto uses hybrid when semantic search is ready, otherwise keyword; each response says which mode ran. Quote a passage only when passage_verified is true.