Alex SaintBack to @alexsssaint

ClipLab

macOS·Apple Silicon·no subscription

Your footage never leaves.

ClipLab reads your master exports straight off disk, transcribes them with whisper.cpp on the GPU, cuts and reframes with ffmpeg, and burns the captions with Remotion. The video stays on the machine that shot it.

  • Freeno price, no plan
  • Open sourcesoon
  • Mac App Storesoon
The ClipLab app: a project list on the left, twenty-four scored clip proposals in the middle with their thumbnails and grades, and the moments list on the right.
ClipLab ranking 24 proposals out of one 8 minute 41 second recording. Each card carries its score out of 100, its hook, flow, value and trend grades, and the reason it was picked.

Instead of OpusClip

What ClipLab is

Point it at a file or a YouTube URL. It transcribes, then makes two passes over the transcript: the first pulls out moments that stand on their own, a complete thought with a setup and a payoff, and snaps every boundary onto a real sentence edge. The second scores each one and keeps the top twenty.

You are not required to trust any of that. The primary mode is a moments file you write yourself, in timecodes, and no model is called at all. What the analysis writes is exactly that file, so you can edit it and re-render without paying for a second opinion.

It ships as a command line tool, as a real Mac app with its own window and Dock icon, and as an MCP server, so a reel can also be built by talking to Claude about it.

It replaces OpusClip, and the upload of your unreleased footage.

Capabilities

What it does

  • Whisper on Metal

    large-v3-turbo, roughly six times realtime on an M-series Mac. Transcripts are cached next to the video and written once.

  • Reframing that follows a face

    A 1.2 MB UltraFace model tracks the speaker so the 9:16 crop moves with them. Without it every crop is a static centre cut, and it says so.

  • Eight caption presets

    hormozi, karaoke, beasty, deepdiver, minimal, clara, editorial, default. Every font is bundled, so a render never touches the network.

  • Scored, not guessed

    Hook 40%, flow 25%, value 25%, trend 10%, with a letter grade per axis and a suggested hook line. The score ranks clips against each other. It is not a view prediction, and the tool says that too.

  • One file per video

    Everything ClipLab knows lives in a project file beside the media. Autosaved, coalesced, committed by atomic rename: 40 parallel writes, zero lost updates.

  • The whole pipeline, offline

    Silence cutting, B-roll overlays, music beds, hook cards and split layouts all run with no network at all. Pass --offline and the analysis runs local too.

Where your material goes

Everything that leaves, named

On this Mac

1 crossing

  • Your video, always
  • Transcription (whisper.cpp)
  • Cutting and reframing (ffmpeg)
  • Face tracking
  • Captions and fonts (Remotion)
  • Music, effects, export

The line

Off this Mac

  • Transcript sentences, and only those

    goes to Anthropic, for ranking clips and writing the post copy

    Only if you switch it on

Three features send text, and only when you have set a key: the clip analysis, the camera director and the post-copy writer. Unset the key, or pass --offline, and nothing leaves at all. No frame of video ever does.

How it works

The pipeline, as built

One analysis, and where the percentage comes from

  1. Download0 to 22%yt-dlp, for a URL
  2. Audio22 to 25%ffmpeg extract
  3. Transcription25 to 55%whisper.cpp, on the GPU
  4. Segmentation55 to 70%moments that stand alone
  5. Scoring70 to 92%hook, flow, value, trend
  6. Proposals92 to 100%thumbnails and previews
Six measurable stages, so the bar never freezes on a long video. Transcription owns thirty points of it because on a real recording that is where the time actually goes.

Requirements

What it runs on

SpecDetail
RequiresNode 20 or newer, pnpm, ffmpeg and whisper-cpp. yt-dlp only for YouTube URLs.
Checkcliplab doctor names anything missing and prints the exact brew command for it.
Modelggml-large-v3-turbo, 1.5 GB, Metal accelerated on Apple Silicon.
Watches~/Movies/ClipLab by default. Drop master exports there and they appear as projects.
Outputout/<video>/: each clip as 1080x1920 H.264 at 30fps, an SRT with the same cue grouping as the burned captions, and a manifest.
Cost, with a keyRoughly 40 to 60 thousand tokens for a 60 minute video, across both passes. Whisper is the bottleneck, not the API.

Status

Open source, and on the store

ClipLab is being opened up. The repository is private today, so there is nothing to clone yet and no public build to download. When it opens it will be at github.com/AlexSSSaint/cliplab, under a licence that lets you read it, build it and keep it.

The Mac App Store version is coming after that, so it installs and updates like anything else on your Mac, without a certificate dance. No date, because a date I have not earned is just a way of being wrong later.

Want it the day it ships?

← All three apps

Surface 10 · Right thigh · sellme.lol ↗