Open-source skill lets Claude Code actually see video frames, all local

Claude can't take video files directly, so pasting a YouTube link normally gets an answer based only on the title or transcript text — the model never sees the actual footage. An open-source command-line tool fixes this gap. Setup involves installing it via pip with a whisper extra, cloning the repository, and copying the skill folder into Claude's local skills directory; after that, pasting a video URL into Claude Code and asking about it works directly.

The tool uses ffmpeg's scene-detection feature to extract only the frames where the picture actually changes, avoiding any extra machine-learning model downloads, and pairs them with a timestamped transcript, handing Claude one folder containing everything plus a manifest file. For slow morphs or animation tutorials, an --adaptive flag scores frames against a rolling neighborhood instead of a fixed threshold. For lectures where slides barely change, a --text-anchors flag forces one frame per subtitle cue.

For long videos, a --grid flag packs multiple frames into contact sheets to avoid overloading the context window. All processing happens on the user's own machine, and the video itself is never uploaded anywhere. It's MIT licensed with 1.6k GitHub stars, and there's a separate paid add-on for creator analytics.