Lexi

Scroll German. Save a word. Lexi makes it stick.

A free, open-source German vocabulary app for English speakers: 10,000+ words from A1 to C2 to learn, and a dictionary of 93,000+ headwords to look things up in. No account, no ads, nothing to cancel — and your progress stays on your device.

Open Lexi → — it runs in the browser; add it to your home screen and it works offline.

Who it is for

English speakers living with German — past the course, somewhere between A2 and B2, who want to read the news and the letter from the Ausländerbehörde without a dictionary open. It works from A1, but Lexi teaches words, not grammar: if you are starting from zero, use it beside a course, not instead of one.

What makes it different

• Nothing to pass. Lexi opens on a German word, not a quiz: one word per screen, with its sound, its meaning and an example, and nothing is graded until you ask. Every few words there is a German news story on topics you choose (tagesschau, SRF, DW, heise), fetched by your own browser — tap any word in it to save it.
• You choose what you learn. Save a word and Üben (practice) brings it back just before you would forget it — FSRS spaced repetition. Cards say why they are in front of you, and recognising a word and producing it (typed, article included) are scheduled separately, because recognising is only half of knowing.
• Checked, not generated. Gender, plural, part of speech and pronunciation come from Wiktionary, never from a model, and a new card is refused outright if any of them disagrees with de.wiktionary or its example sentence does not contain the word.
• Lexi in your ears. Walk turns a session into audio for a walk or a run, 5 to 60 minutes: hear the English, say the German aloud, then hear it. You grade with your headphone buttons; silence grades nothing, and your voice is never scored.
• Your German, as a sky. Every word Lexi teaches is a star, the commonest at the centre, and the ones you know are lit as brightly as you would recall them today. Share it — the picture carries the link.
• Yours. No account, no tracking, no ads. Progress lives in your browser's storage (back it up from Settings). An optional tutor explains sentences on your own AI key, and is off until you add one. The code is MIT; the corpus is CC BY-SA, every source recorded in ATTRIBUTIONS.md.

How it compares

Lexi is small on purpose. What the others do better is real, and worth saying:

If you use…	it is better at…	Lexi is different because…
Duolingo	a whole course — grammar, listening, speaking — as a daily game	it is only vocabulary, for adults, with no game to win or lose, and it goes deep at B1–C1
Anki with a shared deck	anything, in any language, exactly your way	there is no deck to build or debug, and every gender and plural is checked
LingQ, Readlang	reading any text you bring	a word counts as known when the scheduler's evidence says you would recall it, not when you mark it — and it is free

What Lexi deliberately does not do — teach grammar, mark your speech, run leagues, or say "you can now…" from a word count. Ruled 2026-09-05: a drill earns its place if it tests a property of the word (gender, plural, spelling, the German you have to produce) and goes if it tests a rule of the language. Every gloss is in English; other language pairs are an architectural goal, not a shipped feature. docs/VISION.md carries the whole argument, including what it would take to reverse each refusal.

Docs

docs/VISION.md is the anchor — what this is, what it refuses, and what is still undecided. docs/BACKLOG.md is what's next; docs/CHANGELOG.md is what shipped and why. Full index at docs/README.md.

What's inside

Measured against the shipped corpus on 2026-09-26. The two headline counts are guarded by src/lib/publicCopy.test.ts, which fails when the corpus outgrows them.

• 10,000+ word cards across all six CEFR levels, every one with at least two usage examples, German and English, and all but a handful with IPA.
• Cards carry a gloss, gender and plural, synonyms and antonyms, word family and — from B2 — a German-language definition.
• 93,000+ dictionary headwords answer what the cards do not, kept visibly apart: an entry is an unverified Wiktionary gloss you can look up and note, never study.
• 274 topic sectors, rolled up into broad theme groups for browsing and the heatmap.
• FSRS scheduling via ts-fsrs. Every drill mode is its own track, so recognising a word and producing it are scheduled separately — that split is what makes the recall drill honest.
• Local-first: progress lives in IndexedDB with export/import for backup and moving between machines.

Surfaces

Four destinations, and the first one is where the app opens.

• Wörter — the feed. One German word per screen, scrolled: the headword with its article inked by gender, the IPA as a pill you press to hear it, the meaning, and one example sentence. Three actions — ⓘ opens everything the corpus knows, 🔖 saves the word to your next session, and 🎓 practises it now. The feed does not grade you: scrolling is not evidence, so nothing here touches your schedule. The order is the scheduler's, though — what's due, then unseen words from your thinnest topics, then the rest by frequency.
• Themen — what words are there? A search over every card (German or English, umlauts optional), the theme groups with your coverage on each, Decks, the Wortkarte (a semantic map of a sector, with synonym links and node colour by learning status), and Wörter aus einem Text — paste any German you want to read and it says how much of it you can read, which words are in the way, and builds a session out of them.
• Üben — test me. Flip cards interleaved with the three word-fact drills, on one FSRS queue that opens with the words you bookmarked. Interval previews on the grade buttons, German text-to-speech on every string, a hint ladder and near-miss tolerance on typed answers, and a line under each item saying why it is here ("you flipped anbieten a few cards ago — now produce it"). Silence when there is nothing non-obvious to say.
• Fortschritt — how is it going? Dein Deutsch, the sky of every word Lexi teaches with the ones you know lit, a replay of how it filled, and each word's own memory curve; words you know, the two-minute placement test, the A1→C2 level path (which is also the control that decides what you are shown), your goal, the knowledge heatmap, review and recall history, the 7-day due forecast, finished sectors, and blind spots — what you keep getting wrong, ranked by rate, each row a tap into a run of that drill.
• Profile, off the avatar — name, level, streak, goal, topics, flagged cards, and Settings: theme, text size, review intensity (FSRS desired retention), daily pace, which drills are in a session, the HD German voice, and backup / restore.

The loop

browse freely → save what catches you → Üben teaches you what you chose.

Bookmarking is the one thing a feed can honestly record, and it is an instruction to the scheduler rather than a wishlist: buildBriefing serves saved words ahead of the ones it would have picked itself.

Walk — On a walk on the saved-words sheet, or ?walk — is the day's queue as audio for a walk or a run: the English, a cue tone, a pause to say the German aloud, then the German. Only a deliberate headphone press grades — next track knew it, previous track didn't — and a word met for the first time is taught before it is tested. See src/lib/walk.ts.

The three drills

Each is generated from the lexicon and scheduled on its own FSRS track under gym:<mode>:<wordId>.

	Asks	Why it is vocabulary and not grammar
der / die / das	the article of a noun	a German noun without its gender is a half-learned word
Plurals	die Tische, not die Tischen	same
Recall	English in, German out — typed, with the article	the productive half of knowing a word, which recognition never proves

Recall is gated: a word becomes eligible only once its flip card has reached FSRS Review, because asking someone to produce a word they have only just met is a retrieval attempt on something not yet encoded.

Stack

Vite 6 · React 19 · TypeScript · Tailwind CSS v4 · motion · lucide-react · ts-fsrs.

Run

npm install

npm run dev

Then open http://localhost:5173/lexi/ — the base path is /lexi/ locally and / on Vercel.

npm run build      # production bundle to dist/
npm run typecheck  # tsc --noEmit
npm test           # vitest — the pure logic, the shipped corpus, and the public copy
npm run lint       # eslint, including jsx-a11y

Before pushing anything that touches the corpus:

npm run corpus:validate && npm run corpus:selftest

Repository layout

src/            the app — views/, components/ (+ ui/ primitives), lib/, data/
public/data/    the shipped corpus: cards.json, detail.json, sectors.json, freq.json,
                provenance.json, audio.json (vocab.json is canonical, not deployed)
scripts/corpus/     build-time ingestion, audits and one-shot corpus fixes (npm run corpus:*)
scripts/authoring/  the verified card-authoring loop: batch in, machine-gated, audit trail
docs/           VISION (the anchor) · BACKLOG (open) · CHANGELOG (shipped, with reasoning)
design/         logo sources

Nothing under scripts/ ships to the browser: it runs on a maintainer's machine and writes public/data/*.json, which the app fetches at runtime. Corpus JSON is never hand-edited — every change goes through a script so it is reviewable and repeatable.

Data

public/data/ is served as static files and fetched at runtime (see src/data/index.ts → initData) so the corpus isn't parsed inside the JS bundle — the app shell paints immediately and the service worker caches the data for offline reloads. To extend coverage, use the reproducible pipeline in scripts/corpus/: npm run corpus:coverage to see the gap, corpus:build to grow it from open sources.

vocab.json still carries 110 kind: 'grammar' cards. They are filtered out at load in src/data/index.ts rather than cut from the corpus, because the corpus is canonical and reversing the decision should be one line.

Install as an app (PWA)

Lexi ships a web app manifest and a service worker, so it installs on phone and desktop, runs full-screen and works offline after the first load. navigator.storage.persist() runs at boot — without it Safari evicts IndexedDB after about a week of not opening the app, which for a local-first tool is total data loss.

Licence

The code is MIT — see LICENSE.

The corpus (public/data/*.json) is built from Wiktionary/Wiktextract, Tatoeba and the Leipzig Corpora Collection, and carries CC BY-SA 4.0 with attribution — see LICENSE-DATA.md. Every source, its licence, and exactly what is redistributed is recorded in ATTRIBUTIONS.md — if you fork this, that file travels with it. What the app sends where, the credits and the legal notice are on one page: public/legal.html.

Contributing

See CONTRIBUTING.md. The short version: the corpus is never hand-edited, tests and corpus:validate gate every merge, and docs/VISION.md lists what the project deliberately refuses to build.

Notes

Coverage colour uses FSRS state: a card counts as learned once it leaves New and consolidated once it reaches Review. New-card introductions are capped per day, at a pace you can change. Respects prefers-reduced-motion, honours iOS Dynamic Type, and pairs every colour signal with a shape or a label so nothing rides on hue alone.