Claude-real-video (GitHub) is a Python CLI tool that lets any LLM actually “watch” a video by extracting scene-aware, deduplicated frames and a transcript — all locally, no uploading to external services. Unlike fixed-interval frame sampling (e.g., 1 fps) that over-samples static content and misses fast cuts, this tool uses scene-change detection plus a density floor to capture every meaningful visual change while discarding near-duplicates.

Key Features

• Scene-change detection with configurable sensitivity (--scene 0.30)
• Sliding-window dedup (--dedup-window 4, --dedup-threshold 8%) — avoids resending the same shot after a cutaway
• Density floor (--fps-floor 1.0) ensures at least one frame every N seconds
• Hard cap on total frames: --max-frames 150
• Whisper transcription with language detection (--lang auto or specify en, zh)
• Supports URLs (YouTube, Instagram, TikTok, etc.) via yt-dlp and local files
• Outputs a clean folder: crv-out/frames/*.jpg, crv-out/transcript.txt, crv-out/MANIFEST.txt — drop these into Claude, ChatGPT, or Gemini
• Option to generate a visual report of keep/drop decisions (--report)

Installation

pip install claude-real-video           # core (frames + dedup)
pip install "claude-real-video[whisper]"  # + audio transcription

System requirement: ffmpeg (install via brew install ffmpeg, sudo apt install ffmpeg, or winget install Gyan.FFmpeg).

Usage Examples

# YouTube/Instagram link
crv "https://www.youtube.com/watch?v=..."

Local file with English transcript
crv lecture.mp4 -o out --lang en
Frames only, no transcript
crv clip.mp4 --no-transcribe
With cookie file for login-gated content
crv "https://..." --cookies cookies.txt

How It Works

The tool uses yt-dlp to fetch URL-based videos (with optional cookies), then ffmpeg extract passes to grab frames at every scene change plus a density floor. A sliding-window near-duplicate detector collapses repeated shots. Audio is transcribed via the Whisper CLI. Everything stays local — no upload to any cloud service.