Your long video just ranked its own best clips, free

# The Free Clip Ranker Setup

supoclip is a free, open-source repo (AGPL-3.0) that takes one long video and turns it into
ranked, ready-to-post shorts. Paste a YouTube link, and it transcribes the whole thing, finds the
moments worth cutting, scores each one out of 100, crops to vertical following the speaker's face,
and burns in word-synced captions and a hook title. You post the winner first instead of guessing.

The repo (github.com/FujiwaraChoki/supoclip) sits at 1,167 stars and 312 forks as of this writing,
created 2025-07-28, last pushed 2026-09-21, with 29 open issues. It's the same job Opus Clip
charges a monthly plan for, run from your own machine on your own key.

- Transcribe. It reads the audio with AssemblyAI, a local Whisper model, or the video's own
YouTube captions, whichever you point it at.
- Pick. An LLM reads the transcript and pulls out the 3 to 7 segments worth cutting into
clips, not just the loudest ones.
- Score. Every candidate gets four subscores, 0 to 25 each — hook, engagement, value,
shareability — summed to one number out of 100, so you can rank the batch instead of eyeballing it.
- Crop. Face detection drives the 9:16 crop, so the speaker stays in frame instead of getting
cut off mid-sentence.
- Caption. Word-synced captions and an AI-written hook title burn straight into the export.
Pexels B-roll is optional, off by default.

git clone https://github.com/FujiwaraChoki/supoclip.git
cd supoclip
cp .env.example .env
# add your transcription + LLM provider key to .env
docker compose up -d

Open localhost:3000 once the six services are up, paste a link, and watch the scored list fill in.

The software costs nothing and the code is yours to run. It is not a $0 service: you bring your
own API key, and the provider bills you in cents per video, not a monthly plan. Opus Clip's Pro
plan is $29/month for the same kind of ranked output (its Free tier is $0 with a watermark).
supoclip has no watermark and no plan, only the token bill for whichever key you plug in.

Here's the part worth being straight about. The README documents a fully local, no-key path —
local Whisper for transcription, a local Ollama model for the scoring step. We ran it, on our own
9.7-minute video. The transcription half worked well: whisper.cpp (the base model) finished in
7.5 seconds. The scoring half did not. Three small local models — llama3.2:3b, qwen3.5:4b, and
qwen3.6:27b — each returned malformed JSON against supoclip's nested scoring schema, across three
different transcript lengths, and produced zero clips every time. Scoring only worked once we
moved to qwen3:30b-a3b, a 30B mixture-of-experts model with about 3B active parameters — heavier
than what most laptops keep loaded. On that model, real clips came back with real numbers: a
03:16 segment scored 97 (hook 25, engagement 24, value 23, shareability 25, tagged "contrast"), a
04:52 segment scored 93, a 01:12 segment scored 91. So the honest claim is: the software is free
and open, the small-model no-key path as documented did not work for us, and a working run costs
either a bigger local model than the docs suggest or a few cents in API tokens per video.

Comment CLIPS on the video for the setup kit, or grab it straight from the button above. The
kit covers the exact install commands, the four-part scoring schema, the real numbers from our own
test run including which models failed, and the license.

github.com/FujiwaraChoki/supoclip