Our Work

Harvestry

Learning · A Mac app

Stop rewatching. Start learning. Harvestry transcribes your videos, captures the key slides, and exports a polished, interactive study document, entirely on your Mac and entirely in private.

Field
Learning and knowledge work
Platform
macOS 15 or later, Apple Silicon
Technology
WhisperKit, AVFoundation, and Apple Vision
Status
Shipping, version 1.3
The Harvestry app: a library of lectures, a video, and its transcript with captured slides.

Import a Video. Get a Document.

You can rewatch a video, but you can't search it, highlight it, or find the one slide you half remember. Paste a link from YouTube, Vimeo, or a thousand other sites, or drop in a file, and Harvestry turns it into a study page you can read, search, and mark up.

A Harvestry study page open in a browser: the lecture title, a transcript with timestamps, and slides in the margin.

Research-Grade Transcription, On the Device

Every word is transcribed on the Apple Neural Engine with WhisperKit, Argmax's optimized build of OpenAI's Whisper. There's no API key, no upload, and no subscription.

Five model sizes trade speed for accuracy. Whisper's 100 languages are transcribed as they were spoken, or, with Harvestry Pro, translated into English in the same pass.

Harvestry's transcript view, with passages, timestamps, and a captured photograph from the lecture.

Every Slide, Once

Not every frame deserves a screenshot, and the same slide should never be captured twice. Harvestry runs each video through a pipeline that keeps the distinct slides and quietly drops near-duplicate shots of the speaker.

Capture runs on the GPU while transcription runs on the Neural Engine, so a 90-minute video finishes both at the same time.

  1. 1
    Decode

    Frames come off the Mac's hardware video decoder, on the GPU.

    AVFoundation · VideoToolbox
  2. 2
    Find the text

    Each frame is scanned for the shape of a slide: titles, bullets, and blocks of text.

    Apple Vision
  3. 3
    Read it

    Accurate on-device OCR reads the actual words, so every slide is searchable.

    On-device OCR
  4. 4
    Keep it once

    Perceptual fingerprints keep each distinct slide and drop near-duplicates of the speaker.

    Perceptual fingerprints

A Library Worth Coming Back To

Highlights in four colors and notes in the margin are baked into the exported page, visible the moment it opens. Lectures sort into folders and sync through your own iCloud.

The export has no web fonts and no remote scripts, so it opens in any browser, on any device, today or in ten years.

A Harvestry page with a handwritten-style margin note beside a highlighted passage from the U.S. Constitution.

Under the Hood

On-device speech

WhisperKit on the Apple Neural Engine, in five model sizes and 100 languages.

Hardware video decoding

AVFoundation and VideoToolbox pull frames on the GPU, alongside transcription.

Vision and OCR

Apple Vision reads each slide's words; fingerprints keep each slide once.

One self-contained file

Live word highlighting, audio sync, and dark mode in a single HTML page that needs no internet.

Five interface languages

English, German, Italian, Spanish, and French, following the Mac's own setting.

AI only if you ask

Optional notes by Claude or ChatGPT, or with Pro, a local Ollama model that keeps even that step on the Mac.

Private by Design

Transcription, translation, screenshot capture, and export all happen on the Mac, and nothing is sent to us. It's the instinct we bring to patient data: keep it where it belongs.

100languages it can transcribe
5model sizes, from fastest to most accurate
0cloud fees, ever