VideoAgent is an open-source agentic framework on GitHub from the University of Hong Kong’s Data Science Lab (HKUDS) that unifies video understanding, video editing and video remaking in a single LLM-driven system. The project, presented at EMNLP 2026, has gathered about 2,000 stars and is under active development, with repository updates continuing into 2026.
Instead of treating video tasks as isolated models, VideoAgent orchestrates large language model agents that plan, call specialized tools and iterate on results. It can watch and comprehend long videos, answer questions about their content, locate moments, perform edits, and even remake existing footage into new productions. The framework also covers the audio side, with audio understanding and audio editing capabilities, and targets workflows such as podcast production. By exposing these abilities through one coherent agentic layer, VideoAgent aims to be an all-in-one workbench for video intelligence research and application building.
The project is developed by the HKUDS lab, whose other open-source releases are widely used in the data science community, and the repository collects paper details, demos and setup instructions for researchers who want to reproduce or extend the framework.
VideoAgent is free and open source under the MIT license. You can clone it from GitHub and run it with your own models and API keys; there is no paid tier, hosted subscription or usage fee charged by the project itself.
VideoAgent is aimed at AI researchers, video-tool developers and advanced creators who want to experiment with agentic video workflows, from automated video analysis to AI-assisted editing and remaking, using an extensible open-source codebase.