rea-video CLI:
install and first breakdown.

What it is

rea-video is an open-source Python CLI by JuneYaooo, Apache-2.0 licensed. Its brief: reverse-engineer any video, shot by shot — filmed, AI-generated or coded — and plan how to remake it. It runs ffmpeg locally, measures the file, and can call a vision model for the shots the measurements cannot settle.

It is built for agents: you can point Claude Code, Codex or another command-running agent at the repo and a video, and the agent installs the CLI, learns the workflow and runs the breakdown. You can also run the commands yourself.

Install

Requirements: Python 3.10+ and ffmpeg.

git clone https://github.com/JuneYaooo/rea-video.git
cd rea-video
pip install -e ".[ocr]"
rea-video --help
bash scripts/install_skill.sh   # optional: link the agent skill

The first breakdown downloads a roughly 67 MB voice-separation model to ~/.cache/rea-video/models. The project also documents a one-line agent install: ask the agent to install the repo and break down a video.

First breakdown

Run the deterministic, offline pass first:

rea-video decompose ref.mp4 -o out/ref
open out/ref/report.html

That writes a report and a machine-readable breakdown. The project says a 30-second video takes about a minute without a vision model. If you have a video-capable model, add it:

export REA_VLM_BASE_URL=https://.../v1
export REA_VLM_API_KEY=...
export REA_VLM_MODEL=gemini-2.5-pro
rea-video decompose ref.mp4 -o out/ref --vlm

Any OpenAI-compatible multimodal endpoint is accepted; multiple models can be listed as fallbacks. If every VLM call fails, the run does not abort — the report says the VLM step was skipped.

What comes out

• breakdown.json — the video IR: per-shot features, evidence, probabilities, overlays and remake plan.
• report.html and report.md — the human-readable report.
• storyboard.jpg — a contact sheet colored by production method.
• shots/ — per-shot contact sheets, first and last frames.
• packaging/ — subtitle SRT, styled ASS, overlay timeline and sticker captures when OCR is on.
• audio / audio_plan — measured voice, music, effects and the remake plan for sound.

The project also has commands for stems, review tasks, corrections, evaluation and a capsule-cinema remake workflow. The CLI reference in the repo is the source of truth; the summary here is dated 8 Nov 2026.

Optional vision model

The deterministic pass is strong on measurement and weaker on semantics: whether a hand is wrong, whether a surface is physically plausible, whether a shot is a stylized 3D render. A vision model adds a second opinion. The project fuses the deterministic probability with the visual probability and flags conflicts for review. That is more honest than either signal alone.

The project’s published accuracy numbers are from its own evaluation sets. Read eval/RESULTS.md for the sample definitions and reproduction steps before quoting them.

Browser vs CLI

Need	Browser workbench	rea-video CLI
Fast first pass, no install	Yes	No
Cut list, frames, prompt pack	Yes	Yes, plus more
Metadata / AIGC / C2PA	No	Yes
OCR captions and styled text	No	Yes
Voice / music / effects stems	No	Yes
Vision-model review	No	Optional
Runs on a phone	Yes, short clips	No

A practical sequence: use the browser workbench to see the cuts, frames and prompt scaffold. If the question is “how was this made?” rather than “how do I remake the look?”, move the original file to a machine with Python and ffmpeg and run the CLI.

Attribution: the commands and figures above are from the project’s README and CLI documentation, read on 8 Nov 2026. This site does not bundle or redistribute the project’s code.