NeonClipping turns long videos into vertical clips locally
A two-day-old MIT-licensed Python project cuts captioned 9:16 clips from YouTube, Twitch and Kick links using audio signals and a local LLM, with no uploads.
A local clipper, two days old
On 2026-11-21 the repository iamajmeena/neonclipping showed 1,288 stars and 62 forks, collected in the two days since it was created on 2026-11-19. It is an open-source clipping engine: paste a YouTube, Twitch or Kick link, or point it at your own MP4, and it returns vertical 1080x1920 clips with burned-in captions plus a title and a hook for each. The README puts a run at about two minutes. Everything executes on the local machine, there is no account and no subscription, and the README states the server binds to 127.0.0.1 only. The code is Python, MIT licensed, and the repo is tagged with ffmpeg, whisper and yt-dlp among other topics.
How it avoids watching the whole video
Fast mode, the default for videos over 10 minutes, downloads audio only: the README gives 79 MB and roughly 10 seconds for an 85-minute video. For YouTube sources it pulls word-timed captions from YouTube itself; other sources fall back to a local Whisper model. The track is then scored for sudden loudness jumps against a two-minute local baseline, energy in the 2 to 4 kHz band, and quiet-to-loud transitions. The strongest moments are passed to a local LLM through Ollama, which scores each one and writes a title and hook, and the boundaries snap to sentences inside a minimum and maximum length the user sets. Only the chosen sections of video get downloaded, then trimmed and rendered.
Deep mode takes the whole video and adds optical-flow motion detection (OpenCV Farneback) and cut density to the audio signals, shortlists four times as many candidates and gives the model 15 seconds of context on either side of each moment.
Timings in the README were measured on one machine with an RTX 2060 and 16 threads, three clips each: 97 seconds for an 85-minute YouTube stream using YouTube captions, 136 seconds for a three-hour Twitch VOD with no captions using local speech recognition, 21 seconds for a 90-second trailer and 14 seconds for a 90-second local file. The README says results depend on network and CPU.
Ratings retune the ranking
Each clip carries five stars. Ratings are written to data/feedback.jsonl on the user's computer, and after eight of them the weights on the scoring signals are refitted with ridge regression pulled toward the defaults, so a taste for shouting gets ranked differently from a taste for calm explanation. Nothing is uploaded.
Around that core the tool ships 11 layouts, facecam position presets, caption styles named GTA Neon, Impact, Clean, Hormozi and MrBeast, opening headings adjustable from 2 to 60 seconds, 52 caption languages, Hinglish transliteration of Hindi speech, and SRT and WebVTT export. A Live clipping panel records a Twitch or YouTube Live stream, counts chat messages in 10-second buckets, and cuts around spikes that clear more than 2.5 sigma and 2.5x the five-minute baseline, starting about 24 seconds before the spike because chat reacts late.
What it changes for people building apps
The public repository is readable as a pipeline rather than a monolith: app.py holds the FastAPI server and the make-and-export flow, smart.py does transcript scoring and sentence snapping, highlights.py ranks audio and motion, captions_yt.py parses YouTube's word-timed caption format. The ordering is the lesson. Cheap signals filter first, the expensive model only sees a shortlist, and the heavy download happens after the decision. Anyone wiring media processing or model scoring into a web app can copy that shape.
Limits and open questions
The README does not publish a quality benchmark. Its own validation, covering all 11 layouts and five caption styles on one 796.9-second source and a seven-clip restyle in 119.58 seconds, is described as local checks rather than guaranteed runtimes or caption-quality results. The Mac and Linux install.sh has not been tried on a real Mac, per the README. Python 3.10 to 3.13 is required, FFmpeg must be on the path, Ollama with llama3.2:3b is optional but without it clips are picked on audio and motion scores alone, and an NVIDIA GPU is recommended rather than required. Caption language caveats live in a section the source only points to. Only one stream is recorded at a time. There is no mention of a hosted service or an API.
What this suggests for AI app builders
Small local models doing narrow jobs, scoring a moment or writing a hook, keep showing up in tools like this one. If you want the same pattern inside your own product, Draftship is an AI app builder where you describe the app in plain language and it writes, builds and hosts it, with a built-in database, user sign-in and hosting, and paid plans that add model choice, source access, custom domains and GitHub push. The guide to adding AI to your app covers where a model call fits. Draftship has no connection to NeonClipping.
Primary sources
- iamajmeena/neonclipping: Paste a link, get captioned 9:16 clips in ~2 min. Free, local, open-source clipping engine (audio-first, YouTube captions, local LLM). · GitHub trending AI repositories, Nov 21, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.