# Clippy > Clippy is a free, open-source (MIT) AI clipper that runs on your own computer (Windows, macOS, Linux). It turns long videos, podcasts and streams into vertical 9:16 clips for TikTok, YouTube Shorts and Instagram Reels: it picks the moments with the strongest hooks with a local language model (Qwen3-14B via Ollama), frames the speaker with a virtual camera, adds word-by-word captions, cuts dead air, masters the sound and writes a ready post text. Clips are edited by asking in a chat. Nothing is uploaded and there is no subscription. Key facts: - What it is: a local web app with a chat (`clippy start`, http://127.0.0.1:7860) and a `clippy` command line. Paste a video link (YouTube, Twitch, Kick, TikTok, Instagram and other yt-dlp sites) or a file; finished clips appear as cards in the chat. - Output per clip: a 9:16 MP4 with captions, a cover, a title, a description, a comment question, hashtags, a second title for A/B tests and a post text ready to paste. - Editing: requests like "put an opening text on clip 2", "add this song quietly", "make it 16:9", "cut from 'but' to 'anyway'", "undo" become validated edit operations; edits re-render in seconds from a cached base. - Streams: facecam on top and game below; peaks are found by loudness, laughter, motion on screen and the stream chat. "Clip the latest stream of Kai Cenat" finds a streamer's newest finished stream by name. - Daily picks: "every day 3 GTA VI clips" clips the most watched streamer moments of a game's Twitch and Kick category each morning, with a time to post each for a US audience. - Posting: to the user's own YouTube, Instagram and TikTok accounts through their own developer apps, immediately or at a set time; views and watch time are read back and steer later picks. - Languages: every video is transcribed and captioned in its own language (any language Whisper knows); captions can be translated between English and Hungarian. The interface is English. - Pipeline: yt-dlp, Demucs voice isolation, faster-whisper large-v3, MMS_FA word alignment, Qwen3 hook scoring and re-ranking, tight edit on word boundaries, shot detection, YuNet faces and an active-speaker camera, ASS captions, FFmpeg (NVENC when available), two-pass -14 LUFS loudness. - Requirements: Windows 10/11, macOS or Linux; Python 3.12, Ollama, Git, Node.js; about 25 GB of disk; an NVIDIA GPU with 12-16 GB VRAM is recommended. Without one it runs on the CPU or with a free hosted model (Groq, Gemini, OpenRouter). - Install (Windows): `git clone https://github.com/hajdu-patrik/clippy.git`, `cd clippy`, `powershell -ExecutionPolicy Bypass -File install.ps1`, then `clippy start`. macOS and Linux: `sh install.sh`. - Cost and privacy: free and MIT-licensed; videos never leave the computer; account tokens stay local. ## Docs - [Project page with demo video and install guide](https://hajdu-patrik.github.io/clippy/) - [README](https://github.com/hajdu-patrik/clippy/blob/main/README.md): overview, quick start, configuration - [Command line](https://github.com/hajdu-patrik/clippy/blob/main/docs/cli.md): every command and option - [Posting](https://github.com/hajdu-patrik/clippy/blob/main/docs/publishing.md): connecting YouTube, Instagram and TikTok - [Output quality](https://github.com/hajdu-patrik/clippy/blob/main/docs/quality.md): what each stage does, hook scoring - [Local LLM](https://github.com/hajdu-patrik/clippy/blob/main/docs/local-llm.md): Qwen3-14B tuning for a 16 GB GPU ## Optional - [Demo video (4K, one minute)](https://hajdu-patrik.github.io/clippy/assets/demo.mp4) - [Changelog](https://github.com/hajdu-patrik/clippy/blob/main/CHANGELOG.md) - [License (MIT)](https://github.com/hajdu-patrik/clippy/blob/main/LICENSE.md)