Practitioners using language models for video reasoning frequently struggle with context bloat, fixed-interval frame drops, and platform upload constraints. The open-source utility claude-real-video addresses this workflow by preparing video files or online streams locally on your workstation. It isolates visual transitions, prunes redundant frames, extracts spoken audio tracks, and outputs an organized folder containing image files, transcripts, and a descriptive MANIFEST.txt.
The pipeline executes a single chronological ffmpeg pass across the footage, pairing scene-change sensitivity with a configurable density floor to capture rapid edits alongside static sequences. Extracted images undergo downscaled RGB pixel-difference deduplication across a sliding window of recently retained frames, preventing repeated shots in alternating cuts from consuming token budgets. Existing sidecar or embedded subtitles are parsed directly for textual context, falling back to local Whisper transcription when explicit text tracks are absent.
Executing the tool locally keeps source media on your machine, though downstream privacy depends entirely on where you submit the resulting artifacts. Passing these processed frames and summaries to external foundation models still transfers those specific assets to third-party providers. Setup requires Python 3.10 or later alongside system-level installations of ffmpeg and ffprobe. Audio transcription requires the optional openai-whisper package. The repository is released under the MIT license, and users handling gated material must supply their own authorized session cookies.