What is rea-video?
rea-video takes a video file and reports how it was made. It cuts the video into shots, separates voice, music and effects, and for each shot gathers evidence: frame cadence, sensor noise, AIGC tags in the file metadata, garbled text and hands, beat and key. It then classifies each shot as live footage, AI video, avatar, 3D, code-built motion graphics or screen recording, and writes a remake plan split into what to film, what to generate and what to code, with a reverse-engineered prompt per AI shot when a vision model is connected.
It is built to run inside Claude Code, Codex or another agent that can run commands, needs ffmpeg and Python 3.10+, and ships an HTML report with per-shot frames, captions exportable to SRT or ASS, and an audio breakdown. A Gemini or Doubao endpoint is optional. A companion project, capsule-cinema, can generate the AI shots for a remake.
The repository was created on 29 October 2026 by JuneYaooo under Apache-2.0. It credits REA (Reverse Engineer Anything) as inspiration but is a separate project for video, not part of the REA MCP server.
Is rea-video worth it?
rea-video does a job no other open tool here does: it explains how each shot of an existing video was produced and how you would make it again. The report separates filmed shots, AI clips, 3D renders, code-built motion graphics and screen recordings, then turns that into a shooting list, generation prompts and a Remotion approach. It drew 3,494 stars and 490 forks in its first nine days.
The evaluation is more honest than most. On its tuning set of 15 videos the offline tools alone score 98.3% by shot duration; on two holdout sets of 9 and 7 videos, 99.6% and 98.9%. The project's own Chinese results file says the first holdout set is no longer fully independent because rules were added after it failed, and that on 10 re-encoded WeChat videos the offline tools drop to 88.6% until a vision model is added. Those are tens of videos, chosen by the author, not a benchmark.
The catch: "this shot is AI-generated" is a judgement, not a detection. Apart from hard evidence such as AIGC metadata tags or a platform watermark, the verdict comes from heuristics about noise and frame cadence plus, if you connect one, a vision model's opinion, and the agent reviews conflicts by looking at frames. The README's own examples show the failure mode: on a live-action unboxing the offline tool called 35% of the footage AI before agent review, and the vision model called some shots of an AI music video live-action. Nothing here is a trained AI-content detector or a provenance check, so it is not evidence for a takedown, an accusation or a newsroom.
It is also very new. The repository was created on 29 October 2026, there are no tagged releases, and the README says the detailed docs, including the detection rules and remake guide, are in Chinese for now; your agent reads them and answers in your language. Expect the workflow to change week to week.
Our take: treat it as a sharp-eyed colleague who has watched a lot of AI video, good for working out a competitor's production recipe and costing a remake, and wrong often enough that every "AI" label deserves a second look.
Where it earns its place
A performance-marketing lead at a DTC skincare brand
Work out how a competitor's best-performing TikTok ad was made and what a version of ours would cost
Download the original, ask Claude Code to break it down, and read the remake plan: which shots need a phone rig and one set, which can be generated, and how many generations the AI shots need. Offline mode is free; a Gemini vision pass adds API cost.
A motion designer at a small product studio
Rebuild a slick launch video's kinetic type and UI animations
rea-video marks code-built shots and records text, layout and motion timing, which the agent turns into a Remotion project. Live shots and screen recordings are listed as material you supply.
A film-school lecturer teaching AI-assisted production
Show students how a mixed AI and live-action short was assembled
Run the breakdown on a sample, then use report.html with per-shot frames, evidence and lighting notes in class, while pointing out the shots where the tool and the vision model disagree.
Before you commit
- Setup
- Tell your agent to install it from the GitHub URL, or clone and run scripts/install_skill.sh, which installs the CLI and links the skill into ~/.claude, ~/.codex and ~/.agents; needs ffmpeg and Python 3.10+.
- Integrations
- Claude Code, Codex and other command-running agents; optional Gemini or Doubao vision endpoints; capsule-cinema for generating remake shots; Remotion for code-built shots; SRT and ASS caption export.
- Data and privacy
- Offline analysis runs locally; if you add a vision model, frames and audio go to that provider. You supply the video files, and the project asks you not to reuse others' footage, music or likeness.
- Cost at scale
- Free offline; a vision pass and any remake generation are billed per call by Gemini, Doubao or the video model you choose. Not published per video.
- Limits
- Platform recompression removes noise evidence, so downloads need a vision model to stay accurate; sound effects inside voice and music mixes are hard to separate; music source calls were right on 2 of 5 test videos.
What’s good
- Classifies every shot and writes a film, generate or code remake plan
- Offline mode works with no API key or cost
- Published accuracy numbers with their weaknesses stated
- Captions, cuts, BPM and sound effects broken out alongside the visuals
- Runs inside Claude Code or Codex with one install command
What’s not
- AI-or-not verdicts are model and heuristic judgements, not detector output
- Detailed docs are mainly in Chinese
- Nine days old with no tagged releases
- Accuracy drops on platform-recompressed downloads without a vision model
Best for
Video producers and marketers who want to understand how a reference video was made and cost a remake, and who can treat its AI labels as a starting point.
rea-video alternatives
rea-video: common questions
- Is rea-video worth it?
- rea-video is a sharp, unusually candid way to learn how a video was put together, as long as you read its "AI-generated" labels as informed guesses rather than detector results. We rate it 3.7 out of 5.
- How much does rea-video cost?
- Free under Apache-2.0. Tools-only mode runs offline at no cost; a Gemini or Doubao vision model and any video generation for remakes are billed by those providers.
- What are the downsides of rea-video?
- AI-or-not verdicts are model and heuristic judgements, not detector output; detailed docs are mainly in Chinese; nine days old with no tagged releases; accuracy drops on platform-recompressed downloads without a vision model.
- Who is rea-video best for?
- Video producers and marketers who want to understand how a reference video was made and cost a remake, and who can treat its AI labels as a starting point.
- How much does rea-video cost for a team or in production?
- Free offline; a vision pass and any remake generation are billed per call by Gemini, Doubao or the video model you choose. Not published per video.
- Is rea-video safe for company data?
- Offline analysis runs locally; if you add a vision model, frames and audio go to that provider. You supply the video files, and the project asks you not to reuse others' footage, music or likeness.
- What are the best rea-video alternatives?
- The closest alternatives we have reviewed are claude-real-video (crv) and VideoAgent.
Answers checked Nov 7, 2026.