/watch is a Claude Code skill that lets Claude watch a video. You paste a link from YouTube, Instagram, TikTok, Vimeo or about 1,800 other sites, or you point it at a local file, and you ask a question.
It downloads the video with yt-dlp and pulls frames with ffmpeg, up to 100 frames for a long video. The transcript comes from the site captions first. If there are no captions, Whisper runs on your own machine and writes the transcript there. No audio leaves your computer.
The verdict on this one is from the repo, not a reviewer stamp. I checked it today: 144 stars, MIT licence, last updated Oct 24. The README says the same thing the code does. It is a fork of bradautomates/claude-video, with local transcription as the main change.
Cost: free. There is no API key and no account to create. You need Python, ffmpeg and yt-dlp, and the setup script installs what it can.
For me, this is the base of the youtube-watch skill I want to build. It lets me learn from a creator video in minutes, with frames and words together, instead of sitting through the whole thing.