rea-video finds AI-made parts in any video, shot by shot

A marketer looks at a viral TikTok from a rival brand. Often they cannot tell what is a real recording. They cannot tell what came from an image model. A new open-source project aims at that gap.

rea-video is a Python skill for coding agents. It takes any video. It breaks the video down shot by shot. The README says the job is reverse engineering any video. That covers filmed footage, desk recordings, AI videos, and animations. The repository ranks second on Trendshift's daily list. The author says it has gained 1,271 stars in 30 days. The repository was created on 29 October 2026.

What the skill does

The user hands a video file to Claude Code or to OpenAI's Codex. The agent runs rea-video. The tool installs what it needs. It walks through the frames. It writes an HTML report. The report labels each shot as filmed, AI-generated, 3D or coded. It explains how the captions, cuts and audio were made. It ends with a step-by-step recipe to make the same video.

The author names the vision models it can call. These are Google Gemini and ByteDance's Douyin model, known as 豆包 or Doubao. MiniMax H3 is named for generation in the examples. Suno is named as the music source in one case study. All of these are the author's own claims about which models the skill supports.

The accuracy numbers it reports

The README reports its detection accuracy as the author's own measurement. The tool alone labels about 88.6 percent of shots correctly. Adding a vision model lifts that to over 99 percent. These numbers come from the author's own test set.

The author also writes about two limits. Real smartphone screen recordings can be misread as 3D or AI. That fault goes away with the vision-model step. Live-action parts of a video must be re-shot during replication. They cannot be regenerated.

The example on the project page compares a TikTok phone recording of a product. It sits next to an AI-remade version where the product has been swapped. The report flags which shots were replaced.

Where this fits

The skill joins the group of agent-first tools on GitHub. These are published as Claude Code or Codex skills. They are not web products. A user does not open a dashboard. The user runs their agent. The agent points at the video. The agent gives back the report.

The practical readers are content marketers watching a rival brand. They are journalists reviewing a viral clip for signs of editing. They are creators studying how a video they liked was made. All of them already use a coding agent for other work. Adding one more skill is cheap.

The limits are the author's own. The 88.6 percent number is their own measurement. It is from their own test set. The project is a week old. No independent benchmark exists yet.

Source

rea-video on GitHub, by JuneYaooo.