claude-real-video
A Python tool and Claude Code skill that lets an LLM actually watch a video: it extracts scene-aware, deduplicated keyframes and a transcript from a URL or local file, running entirely on the local machine.
claude-real-video lets an LLM go beyond reading a video's transcript by giving it a set of keyframes chosen for how much they actually differ from one another, rather than a fixed frame rate, alongside the audio transcript. In one demonstration, fixed one-frame-per-second sampling produced 58 frames for a 58-second clip, while the tool kept only the 26 that differed and could pack them into a small number of contact sheets, using fewer tokens without losing visual information.
It installs as a Python package with an optional Whisper extra for transcription, and as a skill for Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI and more than fifty other agent hosts, or as a Claude Code plugin. Downloading, frame extraction and transcription all run locally, and the free core tool needs no API key or account.
A separately sold paid add-on extends the same tool with camera-move and cut-rhythm analysis and a timestamped timeline of information the frames alone cannot show, such as gestures and voice pitch. It is written in Python and released under the MIT license.