Local-first, genuinely
Decks, cards and your review history live in your browser's own database. No server holds a copy, because no server is ever sent one.
Spaced repetition that never leaves your device
Donki is a flashcard app for the browser. Cards are written in Markdown, can speak out loud, and are scheduled by the algorithm that has been carrying spaced repetition since the eighties. There is no account to make, because there is nowhere for your decks to go.
Free, installable, and it keeps working with the network off. Or learn with an AI tutor.
Learn with any AI chat
Paste this into ChatGPT, Claude, Gemini or any other assistant. It points the model at Donki's guides: it plans a course with you, teaches and quizzes you, and gives you flashcard decks to import as you go. In chats that can show interactive pages, lessons can speak aloud and let you practise writing characters.
Open https://donki.3sln.com/docs/teaching/ and use that guide to teach me Spanish for a trip to Mexico: plan the course with me first, then teach and quiz me, and give me Donki flashcards as we go.
What you get
Everything below runs in the page you are one click away from. Nothing to install first, nothing to sign up for.
SM-2 scheduling, by session
Cards come back on the classic SuperMemo-2 curve, counted in study sessions rather than calendar days -- so a week away doesn't bury you under a backlog the moment you return.
Four ways to be asked
Self-rated recall, typed answers checked case- and accent-insensitively, multiple choice whose wrong answers come from the rest of your deck, and writing from memory, stroke by stroke, for Chinese, Japanese, Korean and the Latin alphabet.
Cards that speak
A @tts(...) widget reads text aloud in any voice your device
has -- and where it has none for the language, a small neural voice runs
right in the page. It can hide the text for listening practice.
Decks are one portable file
Export a deck and you get a single .donki.json file with the
images and audio embedded in it. Import it anywhere, or share a link that
points at one.
Reminders that know nothing
Opt in and a nudge arrives at the hour you pick. The push carries no payload: your device counts what is due and writes the notification itself.
A card is just Markdown
Write the face the way you would write a note. GitHub-flavoured Markdown, images and audio, plus a couple of extras Donki adds for studying.
- Bold, lists, tables and code, on the face and the answer alike.
- Images and audio referenced by filename and carried inside the deck.
@tts(...)for pronunciation, dictation and listening drills.@stroke(...)for a writing box that checks stroke order as you draw.- A face that is nothing but audio plays itself when the card comes up.
How do you say **hello** in Japanese?
@tts(ใใใซใกใฏ){"lang": "ja-JP"}
{
"type": "deck",
"name": "Japanese Basics",
"urn": "urn:donki:anon:user123:deck456",
"cards": [
{
"urn": "urn:donki:anon:user123:card789",
"face": "How do you say 'Hello'?",
"prompts": [
{ "type": "typed", "answer": "Konnichiwa" }
],
"updatedAt": 1718000000000
}
]
}Your decks are yours to move
Export writes plain JSON -- readable, diffable, and simple enough to generate from a script or a spreadsheet. Import merges by URN and settles conflicts by last write, so re-importing an updated deck edits it in place instead of duplicating it.
Study progress is deliberately left out of the file. A deck you hand someone carries the cards, not your schedule.
Decks an agent can write
This is what the format is shaped for. A deck is one JSON document with a handful of required fields, the cards are Markdown, and every identifier is a string a model can invent โ so "make me a deck for the JLPT N5 verb list" is one response, not an integration.
- Nothing to encode: a deck of text cards carries no attachments at all.
- A text-only model still writes speaking cards โ
@tts(...)is rendered on your device, not shipped as audio. - Multiple choice needs no wrong answers; Donki draws those from the rest of the deck.
- Re-running the agent corrects the deck in place instead of duplicating it, and your review history survives.
- Long decks can be written in parts โ each part uploads, and is regenerated, on its own.
Read https://donki.3sln.com/docs/ and write me
a .donki.json deck of 20 Japanese greetings.
Typed answers for the readings, @tts on every
face so I can hear them, and multiple choice
for the ones that look alike.Where your data actually is
In your browser. Decks and reviews are stored in IndexedDB on the device you study on, and the app works with the network unplugged.
The one thing that ever reaches a server is a push subscription, and only if you switch reminders on: an anonymous endpoint your browser hands out, your time zone, and the hour you picked. Nothing about what you study. Turning reminders off deletes it.
Make your first deck
It opens straight into an empty dashboard. Add a deck, add a card, and you
are studying -- or import a .donki.json file you already have.