HUANGCHIHHUNGLeo/claude-real-video

Most "video + LLM" setups dump every frame into context and burn tokens on near-identical images. This project trims that waste by deduplicating frames at scene boundaries and pairing them with a transcript, so the model gets a coherent temporal picture without a bloated prompt. Works from a YouTube URL or a local file, runs fully on your machine, MIT licensed. The practical use case that stands out: building a tool that lets an LLM answer questions about a recorded meeting, demo video, or tutorial without any cloud upload. It is not a finished SaaS product — it is a script you wire into your own pipeline. Reservation: the deduplication heuristic will miss subtle visual changes in slow-cut content, and transcript quality depends entirely on whichever ASR backend you point it at. But as a foundation for video-aware agents, it is more honest about the problem than most. -> Best for: AI engineer or solo founder building an agent with video context needs.