Shiori（栞, “bookmark”）is a desktop reading companion built around comprehensible input: the primary activity is reading real Japanese text, and every other feature exists to support that. Since 0.2.0, installable language packs take the same reader beyond Japanese — Koine Greek first, with ~19 more you can build from Wiktionary.
Free & open source · Windows x86_64 · no accounts, no subscription, no cloud
Import any book. Read it. Click the words you don’t know — Shiori shows the dictionary entry, the usage register, the conjugation explained piece by piece — and one click later the word is in spaced repetition, anchored to the exact sentence you found it in. The app tracks every word you’ve ever met, grades each book in your library against what you know, and tells you what to read next.
Language support is data-driven: install a language pack, activate it, and the whole app follows — library, reader, dictionary, reviews, statistics, chat. Nothing mixes across languages.
Settings → Languages installs a pack from a folder, a zip, or a URL — with optional SHA-256 verification — imports its bundled texts into the library, and activates the language live, no restart.
Pre-annotated texts derived from MorphGNT carry a lemma, parse, and gloss on every token: interlinear glosses in the furigana slot, parses decoded to prose, stats graded against GNT frequency tiers, and accent-insensitive search that accepts betacode and Greeklish.
Pick from ~19 languages and Shiori compiles the pack locally from kaikki.org’s Wiktextract dump and hermitdave’s frequency lists. Inflection tables become the grammar, frequency is lemmatized into Top-500/1k/2k/5k tiers, and the gigabyte-class downloads resume instead of restarting.
Furigana appears only over words you haven’t learned — and in its strictest mode, only over the first few occurrences of each word per book: scaffolding that fades as you read deeper. Clicking a conjugated verb selects the whole phrase (読んでいる, not 読) and explains the form component by component.
The reading clock is honest, too: pages you flip through too fast don’t count, and the app pauses itself when you wander off.
Chat with a native-speaker persona that converses with you and never corrects you mid-conversation. Instead, your messages come back marked up like a paper: red underlines for grammar errors, orange for phrasing a native wouldn’t use, with the explanation one hover away.
Bring your own brain: Anthropic’s API, any local model through Ollama (nothing leaves your machine), or any OpenAI-compatible endpoint.
Search JMdict by kanji, kana, or any word form. Every kanji in your query gets a card — readings, meanings, school grade, and an animated stroke-order diagram drawn from KanjiVG data, scrubbed stroke by stroke with the scroll wheel — and you can add any hit straight to spaced repetition.
Pack languages get dictionary search too: an inflected query like suis resolves to every candidate lemma — être and suivre — with corpus frequency ranking words within each match tier.
Search Aozora Bunko’s 17,000+ public-domain works and Japanese Wikisource, then import straight into your library — Shift_JIS, ruby markup and all. Aozora’s catalog is cached after the first fetch, so every search after that runs locally and instantly.
Every book is graded against what you actually know, with coverage forecasts, reading velocity, and a JLPT-graded comfortable reading level.
Cards show the word inside the sentence you found it in, framed by its neighbors, plus examples from your other books.
Export your cards with scheduling, or import an existing deck — SM-2 state seeds FSRS.
Unknown / learning / known / ignored, so names and noise never pollute your stats.
The active language with a quick switcher, cards due today with a time estimate from your measured pace, a continue-reading card with time left and a difficulty verdict, and the reading calendar.
Contractions expand in the reader (au = à + le, components clickable), Germanic packs split unknown compounds against their own dictionary, and a candidate picker re-points any ambiguous occurrence.
Pack-defined personas — dead languages disclose a synthetic partner and judge against attested usage — plus composition exercises, translation drills from your own reading, and per-language model overrides.
Dark / light / sepia themes, gothic or mincho Japanese fonts, rebindable shortcuts, adjustable typography.
Everything but LLM calls, online search, and language-pack downloads works with no network — installed packs work fully offline. One SQLite file, one-click backup & restore.
.txt, .md, .html (Aozora), .epub, .pdf — UTF-8 or Shift_JIS, by dialog or drag-and-drop.
Grab the latest shiori-*-windows-x86_64.zip, unzip, and run
shiori.exe. On first launch Shiori downloads its Japanese
reference data (~20 MB) and you’re reading. Other languages
install as packs from Settings → Languages.
