Remove filler words from your video automatically
“Um”, “uh”, “like”, “so”: every unedited recording is full of them. Hunting each one down by hand is the slowest, dullest part of the edit, which is exactly why AgenticCutter starts there: remove filler words first, automatically.
Why does cutting filler words by hand take so long?
Because you have to hear every single spot before you can even cut it. Filler words hide in plain hearing: you listen through the whole recording, mark each occurrence, and nudge the timeline frame by frame until the cut sounds clean. On longer recordings the work grows with the runtime while your concentration drops with every minute. Rush it, and you publish a video that sounds choppier than what was actually said. Do it properly, and the time comes out of your next video.
How does AgenticCutter remove filler words automatically?
It reads instead of listens. AgenticCutter transcribes your recording locally with mlx-whisper, flags every um and filler word directly in the text, and proposes the cut list before anything renders. You read along and decide in plain language via Claude Code or Codex: cut it, cut ums tighter, keep the take. Or you use the native macOS app with its review UI and approve gate. Either way the result is an auto edit of your video that happens on your Mac, not in a browser tab: transcription, cutting, and rendering all run locally, and the cutter only renders once you have approved every cut.
Which filler words does it detect, and what happens to the pauses?
Detection runs on the transcript: whatever mlx-whisper hears as a filler word gets flagged, from ums and uhs to likes and you-knows. For the short grunted ums that never make it into text, um detection from the audio signal can be switched on via the CLI. And the cut does not leave holes behind: the gap a removed word leaves is cleaned up in the same pass, so the sentence closes naturally.
How hard does it cut, and who decides?
As hard as you tell it to, and you decide. Cutting sharpness comes in four levels, from gentle, which only removes the obvious offenders, to strict, which trims everything that does not carry meaning. Every proposed cut sits visibly in the transcript, and nothing renders before your approval: in the app the approve gate is mandatory, in the CLI you confirm the cut list before the render starts.
Does your footage stay on your Mac?
Yes. Transcription runs on-device with mlx-whisper, the cut is executed locally with ffmpeg, and the render lands on your disk as mp4 or ProRes. Your video is never uploaded. If you pick a cloud cut brain, single still frames and transcript text go there. Fully local works with LM Studio.
What about stumbles and retakes?
Stumbles and retakes are a harder problem than ums: the cutter has to recognize that you said the same sentence twice and keep the better take. The detection path for this is built, but the feature status is honest: in progress, not done yet.
Frequently asked questions about removing filler words
Does it catch “like”, “you know”, and other fillers?
Yes. Detection runs on the transcript, so any filler that shows up in text gets flagged. Grunted ums come from the audio signal via the CLI, and every flagged spot is visible before you approve.
Can I keep individual ums?
Yes. “Keep the take” is all the cutter needs to hear. In the native app the approve gate is mandatory anyway: no cut happens without you confirming it.
What happens to the gaps the cuts leave behind?
They are trimmed in the same pass, together with overlong pauses. Visual pause preservation keeps demo pauses where something happens on screen, and the sharpness level controls how tightly everything closes.
Do I need to know how to code?
No. The native macOS app works with drag and drop, a review UI, and an approve gate. If you prefer, you direct the cut in plain language via Claude Code or Codex.
What do you need to run it?
A Mac with Apple Silicon on macOS 26 or newer, plus a Claude or Codex account, or LM Studio for a fully local setup. Export comes out as mp4 or ProRes.
Related features
- Remove filler words and ums
- Adjustable cut aggressiveness
- Transcript-driven editing
Related use case: Generate subtitles