What is codex-video-vision?

Codex Video Vision is a plugin that gives the Codex AI agent the ability to look at and understand video files. Without it, Codex can only work with text, with this plugin installed, you can point Codex at a video and ask it to summarize the content, compare the footage against a script or storyboard, identify missing scenes, or produce a review report. The plugin works by adding a set of video analysis tools that Codex can call. The video_info tool reads basic facts about a file such as its length, resolution, codec, the format the video is compressed in, and whether it has audio. The video_analyze tool goes deeper, detecting scene changes, periods of silence, frozen frames, and motion levels. The video_watch tool pulls out sample frames and an audio transcript so Codex gets an overview of the content. The video_detail tool lets Codex zoom in on a specific moment in the video at higher quality. Transcription of the audio can be handled by Gemini API, OpenAI Whisper API, or a locally installed Whisper model running on your own machine. There are two built-in workflows called skills. The video-perception skill covers general video inspection. The storyboard-review skill is designed for situations like educational video production, where you want to check a finished recording against an intended script or shot list and get a structured report on what was covered, what was missed, and whether the video should be accepted or revised. It is written in TypeScript and requires Node.js 20 or newer, plus ffmpeg and ffprobe installed on your system. It is a port of the MIT licensed claude-video-vision project by Jordan Vasconcelos.