Show HN: claude-real-video feeds videos to any LLM by selecting key frames

Category: Pulse
Status: discussion
Sources: repository, Show HN discussion
Summary: A Show HN reaching 141 points, claude-real-video (MIT) processes a video locally and emits frames, a transcript, and a manifest that any model can read. It selects frames on scene change rather than at fixed intervals, deduplicates near-identical shots by RGB pixel difference with a per-N-seconds density floor, prefers existing subtitles and falls back to Whisper for audio, and accepts URLs via yt-dlp or local files. The stated goal is to send fewer but more relevant frames to cut token usage while preserving comprehension.
Why it matters: It is a token-efficiency pattern for giving text or vision models video context without uploading media to a cloud service.