Harvestry transcribes your videos — lecture
recordings, online courses, tutorials, conference
talks — captures the key screenshots, and exports a
polished, interactive study document — entirely on
your Mac, entirely in private.
You can rewatch a video. You can't search it. You can't
highlight the important parts, annotate a timestamp, or find
that one slide you half-remember without scrubbing back
through everything. Harvestry converts any video — lecture,
tutorial, talk, course — into a structured study document.
Read it, search it, mark it up. Keep it forever.
1 — Import
Drop in any video
Paste a URL from YouTube, Vimeo, or any of a thousand
platforms — or open a local file. Harvestry downloads
and prepares it automatically.
2 — Process
Transcribed and captured — together
Every spoken word is transcribed on the Apple Neural
Engine. Key video image frames are captured
simultaneously on the GPU. Both finish at the same time
— usually in just a few minutes.
3 — Study
Beautiful HTML export
A self-contained study page that opens in any web
browser — on your Mac, iPhone, or any device. Live word
highlighting, text annotations, audio sync, and dark
mode built in. No internet required.
Transcription
Research-grade accuracy. Zero cloud fees.
Harvestry uses WhisperKit — Argmax's optimised
framework for running OpenAI's Whisper on Apple
Silicon. Every word is processed on the
Apple Neural Engine, the dedicated
ML accelerator in every M-series chip. No API key.
No audio upload. No subscription.
Five model sizes let you dial in exactly the right
tradeoff for your video. Tiny finishes a 90-minute
video in minutes. Large Turbo captures technical
vocabulary, proper nouns, and accented speech that
smaller models miss.
Not every frame deserves a screenshot — and the same
slide should never be captured twice. Harvestry runs
a multi-stage pipeline — Apple Vision shape
detection, accurate on-device OCR, and perceptual
fingerprinting — to surface every distinct slide
worth keeping and quietly drop the near-duplicates a
simpler tool would flood you with. Every stage runs
on your Mac.
Powered by Apple Vision.On-device machine learning scans each frame
for text structure. A presentation slide,
whiteboard, map, or labeled diagram
qualifies. A presenter's logo-print shirt, a
news ticker, or a lower-third caption does
not.
Reads every slide — then de-duplicates.Harvestry runs accurate OCR on each
candidate and compares the actual words. A
new headline is a new slide, and it's kept.
The same words are a repeat, and it's
skipped. Two dark slides that look identical
to a pixel comparison are still told apart by
their text.
Knows the slide from the speaker.Talking-head and b-roll shots that differ
only by a gesture are recognized by a
perceptual fingerprint and collapsed to a
single representative — so you get the
slides, not fifty near-identical frames of
the presenter.
Best-frame selection.Catching the transition is only half the
job. Harvestry scans forward from the cut,
scores a short window of frames for
sharpness, and decodes only the winner.
Slides that fade or animate in are captured
fully settled — never mid-entrance.
Manual Curation
The algorithm picks. You perfect it.
Automated capture is thorough, but you know your
video better than any algorithm. Harvestry gives you
a full suite of curation tools so you can take the
result from good to exactly right — without
rerunning any processing.
Full scrubber timeline.Navigate anywhere in the video by dragging.
The current time and total duration are
always visible.
Seek to Clear.Paused on a blurry frame? Press ◀ or ▶ to
step to the nearest sharp, clear frame in
that direction — automatically.
Add or remove screenshots.Capture any frame at full resolution and
insert it into the transcript with one press
— or hover over any included screenshot and
click Remove.
Your Annotated Library
Curate a library worth coming back to.
Every video you process joins your library as a
fully annotated study document — your highlights,
margin notes, and thinking
baked into the HTML, ready whenever
you return. Not a folder of files. A curated
collection that grows more valuable with every
lecture you add.
Margin notes.Select any phrase and write a short note. In
the export it appears in the margin alongside
the amber-underlined phrase — the look of
handwritten marginalia on a printed page,
baked in and visible to anyone who opens the
file.
Highlights and inline notes.Four highlight colours and freeform inline
notes dropped anywhere between passages.
All embedded as standard HTML — visible the
instant the file opens, in any browser.
Your annotations stay with it, always.Every highlight and note stays attached to
its lecture. Open something you processed six
months ago and your thinking is exactly where
you left it.
Organised by you.Group lectures into folders — a whole course,
a speaker's body of work, a topic you're going
deep on. Your entire curriculum in one sidebar,
synced silently across your Macs via
iCloud.
LLM Consolidation
From transcript to insight.
A raw transcript is accurate. It isn't always elegant.
Spoken content is conversational, not written —
presenters repeat themselves, revise mid-sentence, and
circle back. The optional consolidation stage sends your
transcript through an LLM and produces a
second, refined version of your study document:
same facts, same structure, clearer writing.
Standard
Claude API
Routes your transcript through Anthropic's Claude
API. Produces consistently high-quality,
well-structured academic prose. Fast, with no local
setup required beyond an API key.
⚠ Transcript text is sent to Anthropic over HTTPS.
Use local Ollama instead if your video contains
sensitive material.
Pro — Fully Local
Local Ollama
Connects to a locally running Ollama instance. Any
compatible model — Llama, Mistral, Gemma, and
others. Nothing leaves your Mac. No per-token cost.
No internet required.
Requires a capable Mac and a few minutes of Ollama
setup. Once running, it is completely private and
free of any external dependency.
The consolidated document flows into the same HTML export
pipeline as the base transcript — same typography, same
audio sync, same layout. Both versions are fully editable in
Harvestry before export. With Harvestry Pro, you can also
export either version — or both — straight to your Obsidian
vault or any Markdown editor.
HTML Export — Opens in Your Browser
A document worth reading.
The export isn't a raw text dump. Harvestry generates a
polished, fully self-contained HTML page you open directly
in Safari, Chrome, or any browser — styled like a
high-quality editorial publication, with live audio
playback, per-word highlighting, and text markup built in.
Live Word Highlight
As the video plays, each word lights up in real
time. Click any timestamp to jump to that moment.
Slide-Aware Layout
Screenshots are classified at export time. Lecture
slides with readable content render at full page
width — legible without zooming. Presenter footage
floats to the margin as a compact thumbnail, keeping
the text front and centre.
Highlights & Marginalia
Four highlight colours and margin notes are created
in the app and baked into the HTML — visible to
anyone the moment the file opens, no interaction
required.
Offline Forever
No CDN fonts, no remote scripts, no dependencies.
Dark mode and font-size preferences are baked in and
persist across reloads. Works in any browser, on any
device — today or in ten years.
Pro
Obsidian & Markdown export. Harvestry Pro also exports your study document as a Markdown file — YAML frontmatter, per-screenshot assets, and a per-lecture folder structure that drops straight into your Obsidian vault or any Markdown editor. See what's in Pro →
Privacy
Almost everything stays on your Mac.
Transcription, screenshot capture, and HTML export
all run locally. No audio, no video, and no images
are ever transmitted anywhere. The core pipeline has
no network dependencies beyond the video download
itself.
The one optional exception: if
you choose to use the Claude API for LLM
consolidation, your transcript text is sent to
Anthropic's servers over HTTPS. This step is
always opt-in — the base processing pipeline is
fully local whether or not you use it. The Pro
tier supports local Ollama models instead,
keeping even that step entirely on-device.
Transcription — Apple Neural Engine
WhisperKit processes audio entirely on-chip.
No API calls. No uploads. Your spoken words
stay on your device.
Screenshots — GPU, Hardware Decoder &
Apple Vision
AVFoundation extracts frames using
VideoToolbox. Apple Vision scans each
thumbnail for text structure, reads the
actual words via accurate on-device OCR, and
fingerprints every frame so distinct slides
are kept and near-duplicate shots of the
speaker are dropped. No images, no analysis,
and no results ever leave your machine.
Both run simultaneously
Because transcription and capture use
different hardware, they run in parallel. A
90-minute video completes both stages at the
same time.
HTML Export — Fully local
Export generation reads from your disk and
writes to your disk. The only network call
is the optional yt-dlp video fetch.
Localisation
Available in five languages.
Every label, prompt, status message, and tooltip is fully
localised. Harvestry respects your macOS language preference
automatically — no configuration required.
🇬🇧 English🇩🇪 German🇮🇹 Italian🇪🇸 Spanish🇫🇷 French
Ready to learn more from everything you watch?
Download Harvestry and turn your first video into a study
document — free. No subscription, no account. Drop in a file
or paste a URL and press Begin Processing.
