rea-video

rea-video is an open-source Python CLI built for coding agents (Claude Code, Codex, and anything that can run shell). Feed it a local video and it splits shots, separates voice/music/effects, gathers frame and audio evidence, then decides per shot whether the picture was live footage, AI video, avatar, 3D, code/template, screen recording, or a mix. You get a one-line mix summary, a remake plan (film vs generate vs code), and a full HTML report with frames, prompts, lighting notes, and caption timing. Optional vision models (Gemini or Doubao) raise accuracy on platform-recompressed downloads; tools-only mode stays offline and free on original files.

Founders copying viral product clips still guess camera setups, prompts, and cuts by eye. rea-video turns that into evidence and a recipe: what was live, what was generated, what was motion graphics, and how to rebuild it with your own assets instead of a vague "make something like this" prompt.

Install ffmpeg and Python 3.10+, then ask your agent to install the repo and break down a path or download. The pipeline is split → gather evidence → decide with hard signals first → plan remakes batched for real video-model limits. Captions export to SRT/ASS. Pair with the author's cheap-claude-video / capsule-cinema flow if you want AI shots regenerated on the original timeline; live shots stay a shooting plan, code shots become a Remotion project.

Not REA (apps and binaries) and not a one-click lookalike renderer like Hypit. The product is understanding production: live vs AI vs code per shot, then a cheap remake recipe. Hypit ships a new video; rea-video ships the forensic brief those tools skip.

• Public demo available

• REA ↗
• Craft Anything ↗
• OpenMontage ↗