A collection of tools for AI-powered video editing, built toward a fully agentic video pipeline.
The goal is to build a set of tools that an AI agent can call to edit videos end to end — picking the best clips, analyzing styles, burning captions, and more. Each tool in this repo is a standalone module that does one specific thing.
Current state: The tools work independently as hardcoded flows. Full agentic orchestration (where an AI decides which tools to call and in what order) is the next step.
A web app that downloads a YouTube video, shows an interactive timeline with synced captions, and uses AI to find the best clips for YouTube Shorts.
- Paste a YouTube URL, choose quality, the video downloads automatically
- Interactive timeline with draggable in/out point selection
- AI pipeline (Cerebras LLM + semantic embeddings) scores every segment for viral potential
- One-click export of trimmed clip + captions
Upload any video and get back a full JSON breakdown of how its captions are styled. Built for reverse-engineering viral short-form content.
- Detects font, color, stroke, shadow, and position
- Detects entry/exit animations and karaoke-style word highlighting
- Counts cuts per second and zoom events
- Filters out watermarks and brand overlays
Older standalone utilities used before the above tools were built:
- YtVideoDownloader — Download YouTube videos via a simple Streamlit UI
- YtCaptionDownloader — Download captions from YouTube in SRT, VTT, TXT, or JSON format
- ClipSelector — Earlier version of the clip scoring logic
Each tool has its own setup. See the README inside each folder.
Quick links:
- Python 3.9+
- ffmpeg installed and in PATH
- Node.js 18+ (for VideoSelection frontend only)
- Wrap each tool as a callable function with a standard interface
- Build an AI agent layer that chains the tools together
- Add a B-roll search tool
- Add an auto-caption burner tool
- Full end-to-end pipeline: YouTube URL in, finished short-form video out