NeonClipping: A Free AI Clipping Tool That Turns Long Videos Into Captioned Vertical Shorts on Your Own PC
The trick behind fast AI clippers is that they barely look at your video. Now that trick runs free on a gaming PC from 2019. Here is what it gets right, and the moments it will never find for you.
The fastest AI clipping tools share a secret that sounds like cheating. They don't watch your video.
They listen to it. A shout, a burst of laughter, a sudden jump in volume after two quiet minutes: those spikes show up in the audio long before anything visual would need analysing. Pair that with captions the platform already made, hand the transcript of each spike to a small language model, and you have most of what a paid clipping service does. The expensive part, decoding hours of video frame by frame, mostly never happens.
NeonClipping is that trick, packaged as a free, open-source app that runs entirely on your own computer. It went up on GitHub on November 19 and passed roughly 790 stars in its first day, which for a tool built by one person is a loud signal. Its maker, Ajay K Meena, is a director, cinematographer, colorist and DaVinci Resolve educator in Jaipur. That background shows in the details.
My position after reading its documentation closely: if you clip your own long videos (podcasts, streams, interviews, lectures), a local tool like this is now good enough to do the first pass that people pay monthly for. It is not good enough to replace your judgment, and it tells you so itself.
What it does in two minutes
You paste a link to a long YouTube, Twitch or Kick video, or point it at your own MP4. You set a clip length range, say 30 to 80 seconds. You press Make clips. A few minutes later you have vertical 1080x1920 clips with burned-in captions, plus a title and a hook line for each one.
The author published timings from one machine with an RTX 2060 graphics card, three clips each time:
- An 85-minute YouTube stream that had YouTube captions: 97 seconds.
- A three-hour Twitch recording with no captions, so the app had to transcribe speech itself: 136 seconds.
- A 90-second local MP4: 14 seconds.
These are the author's own numbers, and the README says plainly that yours will depend on your connection and processor. Still, the shape is the point. A long video produces clips in roughly the time it takes to make coffee, on a card most gamers retired years ago.
Nothing is uploaded anywhere. By default the app's server only listens on your own machine; an optional phone-remote mode, off unless you start it, opens it to your home Wi-Fi behind a PIN. There is no account and no subscription. It is MIT-licensed, which means you can use it for paid client work, modify it, and redistribute it.
How it finds the good bits
This is the part worth understanding, because it tells you both why it works and where it fails.
In the default Fast mode, NeonClipping downloads only the audio first. For an 85-minute video that is about 79 MB and roughly ten seconds. It then takes YouTube's own word-timed captions instead of running a speech model, which is faster and, by the author's account, more accurate on noisy game audio. For sources without captions, it falls back to a local speech recogniser.
Then it scores the whole soundtrack for a handful of signals:
- Loudness jumps compared with the previous two minutes.
- Raised voices, measured as energy in the frequency band where shouting and screaming live.
- Phrases that tend to mark a moment ("oh my god," "run," "police," plus Hindi and Hinglish equivalents like "bhaago" and "arre yaar").
The best-scoring moments go to a small language model running on your PC through Ollama, a free app for running such models locally. The model reads each moment's transcript, scores it for emotion, punchline and payoff, and writes a title and a hook. Start and end points snap to sentence boundaries so clips don't begin mid-word.
Only then does the app download the chosen sections of video, not the whole file, and render them.
A slower Deep mode adds visual signals: fast motion, flashes, red-and-blue light flicker, and how often the source cuts. It downloads the whole video first, so it is slow on long sources by design.
The feature that matters most: it learns your taste
Every clip gets a five-star rating box. After you rate eight clips, NeonClipping re-weights its signals toward what you liked. Your ratings stay in a file on your computer.
This is the honest answer to the big weakness of every automatic clipper. "Good moment" means different things on a gaming stream, a cooking podcast and a design talk. A shout-heavy scoring system will love a streamer who yells and will ignore a host whose best line is delivered in a near-whisper. Rating teaches it which kind of channel you run.
Treat those first eight ratings seriously. Rate the clips you would actually post, not the ones that are merely competent.
The finishing options are a colorist's list
The finishing controls read like they were written by someone who has actually delivered vertical video to clients:
- Eleven layouts for streamer footage: camera above or below gameplay, blurred fill, full-screen crop, corner overlays, a circle facecam and side-by-side.
- A Find my facecam button that looks for a rectangle that stays put and contains a face, so the game's minimap doesn't get mistaken for your webcam.
- Five caption styles, editable word by word, with SRT export so you can take captions into Resolve or Premiere.
- Opening headline text with adjustable length, plus templates that wrap a clip in a fake post card, a hook box or a news-style frame.
- Logo, credit line, an outro ribbon and a music bed that rises at the end.
- Hindi speech written in English letters (Hinglish) by default, which is how a large audience actually reads captions.
You can also edit a saved clip's layout and captions afterwards, and apply one change to every clip from the same video, without re-transcribing anything.
There is a live mode too. Point it at a Twitch channel or a YouTube Live link and it records the stream, watches chat speed, and cuts a clip around each chat spike. The author tested it on Twitch; the README says an automatically clipped YouTube stream was not run end to end.
Put this into practice
Here is the test I would run before trusting it with anything that matters.
1. Pick a video you have already clipped by hand. Your own podcast episode or stream, where you know which three moments you posted and how they performed. This gives you an answer key.
2. Install it. On Windows, the easiest route is the setup file on the project's Releases page, which installs Python and FFmpeg (the free video toolkit it depends on) if they are missing. On Mac or Linux, clone the repository and run the start script; the author says the Mac installer has not yet been tested on a real Mac. Then install Ollama and pull the small model the README names (llama3.2:3b). Without it, clips are picked by the audio and motion score alone, with no titles or hooks.
3. Run Fast mode with your usual clip length. Set the range you actually post, not the default. Choose the facecam position or draw the boxes yourself; the README is blunt that a wrong preset will show the wrong part of the frame.
4. Compare its picks with yours. Did it find your best moment? Did it find one you missed? Did it pick three loud moments that were loud for no reason?
5. Rate everything, then run it again on a second video. Eight ratings is the threshold. The second run tells you whether its sense of your channel improved.
6. Export captions as SRT and finish in your editor if you care about the look. Burned-in captions are fine for volume. For flagship clips, the SRT gives you editable captions in your own house style.
If it finds two of your three hand-picked moments on the first run, it has earned a place in your weekly routine as the first pass. If it finds none, your channel's best moments are probably quiet ones, and the next section explains why.
Where it breaks
The README has an Honest limits section, which is rarer than it should be. Here it is in plain terms, with a few of my own additions.
It finds loud, not good. The author's own words: it finds "loud, energetic, quotable moments, not guaranteed 'viral' ones." A perfectly timed dry joke, a long pause before a reveal, a guest who tears up without a sound: these have no audio spike. You will have to find them yourself.
The language model cannot see the screen. It reads text only. A visual gag with no dialogue is invisible to it. Deep mode's motion signals help, but they are simple measurements on a tiny low-resolution copy. They rank moments; they do not recognise what is happening.
Captions are only as good as the source. With YouTube links it uses YouTube's own automatic captions, which are imperfect, especially for Hinglish. Read every caption before posting.
No captions means slow on a CPU. Twitch recordings and your own files need speech recognition. With an NVIDIA graphics card that is fast; without one, expect to wait.
Deep mode is lightly tested. The author verified it on a short video only.
It is one day old. One person maintains it. Expect rough edges and changes. That is the trade for free.
Rights are still your problem. This is my caution, not the README's. The tool will happily download and clip anyone's video. Clipping your own channel is the clean case. If you clip for a streamer or a client, get permission in writing and check the platform's terms.
Who should switch, and who shouldn't
If you post clips from your own long-form content every week, try this first. The first-pass job (find candidates, crop to vertical, add captions) is exactly what it does well, and doing it locally means no per-minute bill and no uploading unreleased episodes to someone else's server.
If your content's best moments are quiet, visual or slow-building, keep doing the selection yourself and use NeonClipping only for the finishing: layouts, captions and batch re-styling. That part alone saves real time.
If you run an agency clipping dozens of channels with brand-specific templates and review workflows, a paid platform's team features still matter. A one-person, one-day-old project is not that, yet.
The subscription clipping apps sold access to a technique. The technique is now free and documented in a README. What they can still sell is polish, support and taste, and taste is the one part you were always going to have to bring yourself.
Run it on one episode you already know by heart. The gap between its picks and yours is the most useful thing you will learn about your own channel this week.