AutoClip is an open-source video clipping tool that turns long recordings into short highlight clips by reading the subtitles instead of watching the pixels. It is built for people who produce talk-heavy footage and then have to mine it for shareable moments: podcasters, interviewers, streamers with hours of replay, and instructors sitting on recorded lectures. The project is written mainly in Python, ships under the MIT license, and has gathered around 8,400 stars and 1,500 forks since it first appeared in July 2025.
What it does
Point AutoClip at a long video and it performs the first pass a human editor would otherwise do by hand:
- analyses the subtitle track to locate the moments that stand out
- scores each candidate moment, so the strongest ones surface first
- writes a title for every clip it proposes
- cuts the individual clips and assembles a combined highlight reel
- applies export presets that burn captions for vertical formats such as Shorts, Reels and Douyin
The input is either a file you drop into the import area — with an SRT subtitle file alongside it if you already have one — or a YouTube link you paste. What comes out is a set of finished clips plus a compilation, ready to upload rather than ready to keep editing. The channel's demo runs a single podcast episode through it and ends up with roughly a dozen postable clips.
How it works
Because the selection step works on text, the expensive part of the job is a language-model pass over a transcript rather than any kind of frame-by-frame vision analysis. The model marks the spans worth keeping, ranks them and names them; after that, cutting the timecodes and burning captions are ordinary video-processing steps. That design is also why the tool suits interviews, podcasts, courses and live replays specifically — the README names exactly those four cases.
AutoClip is not tied to a hosted model provider. It can run against your own local models, which keeps the footage and the transcripts on your machine and removes the per-minute API bill; the video in the catalog makes the point that there is no cloud lock-in.
The same pipeline is reachable three different ways: a packaged desktop application, a web interface you run yourself through Docker, and a command line that is also exposed over MCP, so an assistant or agent can trigger a clipping job as a tool call instead of a person clicking through a UI.
Getting started
Desktop installers are published on the releases page for macOS on Apple Silicon and for Windows on x64, which is the shortest path if you just want to see it work. If you would rather self-host, the Docker web interface gives you the same import screen in a browser, and the CLI covers batch or scripted use. The repository carries an installation guide that walks through the first export end to end, plus a separate troubleshooting document, and the project has its own site linked from the README.
One practical note for an English-speaking audience: the repository's front page is written in Chinese, but translated READMEs are provided, and since version 1.3.1 the application interface and the project website are available in English along with several other languages, with the language selector in the top bar or following the system setting. Your own footage and anything the tool generates from it stay in their original language.
When to use it / when not
Use it when the value of your video is in what is being said. A two-hour interview, a lecture recording, a stream replay where three good exchanges are buried somewhere in the middle — that is precisely the shape of problem this attacks, and the fact that it produces a title per clip means the output is closer to publishable than a bare list of timestamps would be.
Skip it when there is no dialogue to analyse. Music sets, sports footage, silent b-roll and anything whose best moments are visual give the subtitle-driven scoring nothing to work with. It is also not a general-purpose editor: it makes cuts and burns captions, but it will not replace a timeline you use for colour, sound design or multi-camera work. And if your desktop is Linux, note that the installers advertised in the README cover macOS and Windows, so the Docker route is the one to plan for.
Anyone who publishes long-form talking video and currently pays an editor, a subscription clipping service, or their own evenings to cut it down should look at this seriously. It is free, MIT-licensed, runs locally if you want it to, and automates the genuinely tedious part — finding and naming the good bits — rather than just wrapping a trim tool in a nicer interface. The honest limit is its scope: it is a highlight machine for spoken content, and outside that lane it has little to offer.