Hours of gameplay in.
Shorts out.
shardreel finds the peak moments in long recordings and cuts them into captioned 9:16 clips for Shorts, Reels and TikTok. Transcription, analysis and encoding all run on your Mac.
$ cd web && bun install && bun run build $ cargo run --release clipforge listening on http://127.0.0.1:8787
The app
A local web UI with live progress.
Drop in a file or paste a URL, watch each stage run, then review the clips in a grid. Everything is served from one process on your machine.
gta-session-evening.mp4
/Users/you/Movies/gta-session-evening.mp4
Cops chase ends in a crash
#gta6 #chase #shorts
1:12 - 1:42
Five stars in thirty seconds
#gta6 #wanted #gaming
14:03 - 14:33
The truck did not stop
#gta #crash #fail
31:48 - 32:20
Interface preview. Titles and footage are illustrative.
How it works
Six stages, cached after the first run.
Each stage saves its result, so re-rendering with a different layout skips transcription and analysis.
01
Fetch
A local file, an upload, or a YouTube or Twitch URL through yt-dlp, with a live download bar.
02
Transcribe
Whisper large-v3-turbo on the Apple GPU through Metal, with word-level timestamps.
03
Analyse
Per-second loudness spikes and on-screen action from scene-change scores.
04
Pick
Peaks become windows with build-up before and payoff after, snapped to sentences and cuts.
05
Rank
Optional. A model re-ranks within guard rails and writes the title, hook and hashtags.
06
Render
1080x1920 H.264 or HEVC, loudness-normalised, several clips in parallel.
Layouts
Four ways to fit 16:9 into 9:16.
Pick a layout per job and see it on a real frame of your video before you render.
Blur
Zoomed gameplay over a blurred copy of itself.
Crop
A full-height 9:16 slice. You choose the horizontal position.
FacecamBeta
Your webcam on top, gameplay below. Draw the box on a real frame.
Fit
The whole 16:9 frame, no zoom. Good for cutscenes.
Features
Built for gameplay, not talking heads.
Native on Apple Silicon
Whisper runs on the GPU through Metal. Encoding uses the media engine through VideoToolbox. Nothing leaves your Mac unless you turn on an AI model.
Finds the peak
Loudness spikes, on-screen action, speech pace and hype words combine into one excitement curve. Silence never wins.
Captions that follow the voice
Word-by-word highlighted captions and a hook banner, burned in with libass from real word timestamps.
AI ranking, optional
Gemini, OpenRouter, LM Studio or Ollama can re-rank candidates and write titles. If the model fails, the run finishes on signals alone.
Re-render in seconds
Analysis is cached per video. Change the layout, captions or clip count without transcribing again.
Yours to keep
No accounts and no subscription. Clips are saved to a folder you choose, ready to upload by hand.
Point it at a recording.
Setup takes a few commands. The guides walk through each one.