# Clips Studio

> Free, open-source AI video clipping that runs entirely on your own Windows PC.
> Paste a Twitch VOD, Kick VOD or YouTube link and it finds the moments worth
> posting, crops them to vertical with the speaker kept in frame, burns in
> word-synced captions and writes the titles. No uploads, no subscription, no
> per-clip fee. AGPL-3.0.

Clips Studio is a desktop application, not a web service. Everything — the
transcription, the clip selection, the subject tracking and the language model
— runs on the user's own machine. It is a local-first alternative to cloud
clipping tools such as OpusClip, aimed at creators who would rather not pay per
video or upload their footage to someone else's server.

## What it actually does

- **Transcribes** with faster-whisper at word level, so captions land on the
  right syllable and clips start on real sentence boundaries.
- **Scores** candidate moments with a local language model reading the
  transcript alongside audio and visual signals — loudness, laughter-shaped
  bursts, scene cuts, motion, on-screen reactions.
- **Reframes** to 9:16 with YOLO pose tracking and active-speaker detection, so
  the person talking stays in shot and it cuts between speakers rather than
  zooming out to hold both.
- **Edits**: a built-in editor for fixing anything the AI got wrong, caption
  styling, watermarks, multilingual captions and dubbing.

## What it is good and bad at

- **Good at:** IRL, just chatting, podcasts, vlogs, interviews. That is what it
  is tuned for and tested on.
- **Not yet:** gaming, split-screen and reaction videos. The framing did not
  reliably find the action, and subject tracking mistook characters inside a
  game for the streamer, so it is held back rather than shipped half-working.

## Requirements

Windows 10 or 11, 64-bit. 16 GB RAM recommended, ~20 GB free disk. An NVIDIA
GPU is strongly recommended — it runs on CPU, just far slower. One installer
carries the app, the engine, FFmpeg, the Ollama runtime and the tracking and
transcription models; the language model downloads on first launch, sized to
the graphics card.

## Links

- [Website](https://colingpt9.github.io/clips-studio/)
- [Source on GitHub](https://github.com/ColinGPT9/clips-studio)
- [Download](https://github.com/ColinGPT9/clips-studio/releases)
- [Architecture](https://github.com/ColinGPT9/clips-studio/blob/main/ARCHITECTURE.md)
- [Changelog](https://colingpt9.github.io/clips-studio/changelog.html)
- [Roadmap](https://colingpt9.github.io/clips-studio/roadmap.html)
- [Known issues](https://github.com/ColinGPT9/clips-studio/blob/main/KNOWN-ISSUES.md)
- [Clipping Twitch](https://colingpt9.github.io/clips-studio/twitch.html)
- [Clipping Kick](https://colingpt9.github.io/clips-studio/kick.html)
- [Clipping YouTube](https://colingpt9.github.io/clips-studio/youtube.html)
- [Local vs cloud clipping](https://colingpt9.github.io/clips-studio/local-vs-cloud.html)

## Status

Alpha. It works and is used daily, but it is early and has rough edges, which
are listed honestly in the known-issues document rather than discovered by
users.