wasi-master/wmkeyboard

Private. Offline. Yours.

An open source multilingual Android keyboard with prediction, glide typing, transliteration, emoji search and a 75+ tool toolbox that all run on the phone and never phone home.

Install · Tour · Languages · Privacy · Editions · FAQ · Build · Read the docs

WM.Keyboard.-.Private.Offline.Android.Keyboard.Official.Trailer.mp4

A keyboard is the one app that sees everything you write. Your messages, your passwords, your half-typed thoughts you deleted before sending. Somewhere along the way the popular keyboards decided that was a business model. The prediction got better, and the price for that was a copy of your keystrokes getting sent to somebody else's server.

The keyboards that refused that deal were mostly forks of the same decade-old Android code. Good bones, honest intentions, but no glide typing without a proprietary blob, no real toolbox, and nothing for the language I actually type in. Bangla was an afterthought everywhere: either a phonetic engine bolted on with no prediction, or a layout with no smart features behind it.

So I wanted a keyboard that did three things at once. Type Bangla the way Avro taught a whole country to type it. Carry the features I reach for all day, clipboard, snippets, translate, calculator, scanner, without leaving the text field. And keep all of that on the phone, where it belongs. This is that keyboard. It is written from scratch in Kotlin and Jetpack Compose, keyboard surface included, and the whole intelligence layer runs offline.

Type a word the way it sounds in Latin letters, and the keyboard finds the word you meant. Every one of these runs on the device, with no network and no delay.

The same prediction engine sits behind every one of them, so glide typing, autocorrect and next-word suggestions work in Bangla exactly as they do in English.

The features the keyboard has

• Prediction that learns, locally. A trie-lattice beam decoder shared by tapping and glide, n-gram context ranking, and on-device learning of your words, bigrams and trigrams. One tap clears all of it.
• Autocorrect you can argue with. Corrections are gated by a likelihood ratio rather than a blunt threshold, an undo chip appears after each one, and a word you fix by hand is learned.
• Glide typing in any language with a word list, no proprietary library. English and Bangla ship with lists built in. Downloads cover 330+ more.
• Gestures everywhere. Spacebar as a trackpad, swipe backspace to delete words, swipe space to switch language, drag a modifier for chords, peek a layer by dragging off shift, and much more!
• Modes. One-handed, split, floating, and per-app modes that swap the whole keyboard, toolbar included, when a particular app opens.
• Field awareness. Enter, the number row, suggestions and learning all adapt to the field: passwords, code editors, terminals, URL bars, one-time-code fields.
• Hardware keyboards get a hotkey layer, hint overlays and tool shortcuts.

Typing basics · Glide · Gestures · Modes · Autocorrect

• 826 languages, 1,700+ layouts from one data-driven registry: native scripts, InScript variants, romanised entries, CJK, minority Cyrillic alphabets, and a dozen constructed ones. A fresh install starts in whatever your phone is already set to.
• Bangla, done properly. Avro phonetic, Probhat with aspirates on shift, Jatiya, Khipro Classic and Dark, conjunct-aware backspace, and Bangla-English mixing in one field.
• CJK input with composers for pinyin, jyutping, kana and hangul, plus Vietnamese Telex.
• Layouts you can edit, build from JSON, or import from FlorisBoard, HeliBoard, FUTO, and Keyman.
• Odd ones too: musical notation, braille chording, Morse, fancy-text alphabets, and a T9 keypad for languages that had one.

Languages overview · Bengali · CJK · Custom layouts

• Semantic emoji search in 126 languages. party finds 🎉 🥳 🍾 🎂, and so does the word for party in other languages.
• Emoji 17.0, per-person skin tones, kaomoji, and long-press to send Google's animated emoji.
• Sticker packs you make from your own photos, with a background cutout that runs on device. WhatsApp and Signal packs import directly.
• GIF search when you want it, off when you don't.

Emoji picker · Stickers

77 tools sit one tap behind the toolbar, and you drag the ones you use onto it. Most of them work in airplane mode.

The toolbar · Clipboard · Whisper · AI tools

• A full theme editor: palettes, fonts per script, key shapes, textures, decals, particles, photo backgrounds, per-key styles, with a live preview while you edit.
• Material You, AMOLED, and 29 built-in themes across 11 families, including Dracula, Nord, Solarized, Catppuccin and Tokyo Night.
• Icon packs, sound packs, and themes that export as files you can share.

Themes · The editor

The typing engine has no network code in it. Prediction, autocorrect, glide, layouts and learning behave identically with the radio off.

• No analytics SDK, no crash reporting, no account, and no server of mine for the app to reach.
• Learned words, clipboard history and typing stats live in app storage and wipe in one tap.
• Password and secure fields turn off learning, suggestions and clipboard capture on their own. Incognito mode does the same on demand, and follows private browsing automatically.
• Every tool that touches the network says so on its own settings page and stays off until you open it. A network activity log shows every request the keyboard has made.
• Backups can be encrypted with a passphrase and written wherever you point them.
• The one exception worth naming: handwriting and the scanners in the Full edition are built on Google's ML Kit, which reports its own diagnostics to Google. The Lite edition carries no Google library at all, and a build flag can remove the internet permission entirely.

Privacy at a glance · Network policy · Your data on device

• Addon repositories serve 14 kinds of addon: themes, layouts, dictionaries, icon packs, sound packs, sticker packs and more. Host your own with a static JSON file.
• Lua plugins run sandboxed inside the keyboard and can add whole new tools, with a permission model and an in-app IDE.
• Automation intents let Tasker and friends switch layouts, modes and languages.

Addons · Plugins · Repository format

One codebase, two APKs. Everything about typing, prediction, themes, languages and privacy is identical. What differs is which on-device machine learning ships in the box.

Sizes are from 0.5.12 and drift with dependencies. The UI-language choice only affects the words the settings app shows you. Every layout and dictionary is in both.

Full vs Lite

• Google Play carries the Full edition. Play delivers the heavier on-device parts (the local LLM runtime, the translation engine, handwriting, the offline-voice interpreter) the first time you use them, so the install stays small, and updates it like any other app.
• F-Droid carries the Lite edition, built from source and checked byte-for-byte against the -fdroid.apk on the matching GitHub release.
• GitHub releases carry both editions, each in English-only and all-languages builds, for arm64-v8a, armeabi-v7a, x86_64 and a universal APK. Almost every phone made since 2017 wants arm64-v8a. A SHA256SUMS.txt sits next to them.
• Obtainium tracks the GitHub releases and updates you automatically. The release APKs also carry their own update check, which asks GitHub and shows a card on the Home screen.

Install it, open the app, follow the setup card. Android will warn you that a third-party keyboard can see what you type. It says that about every keyboard, Gboard included, because that is what a keyboard is. F-Droid and GitHub builds share a signing key from 0.5.10 on, so moving between them is an ordinary update. Google Play signs its copy with its own key, so moving between Play and the other two takes a backup, an uninstall and a restore.

Needs Android 7.0 or newer.

Installation · Setup · A two-minute tour

The docs have a longer FAQ and a troubleshooting page.

wmkeyboard.pages.dev is the full manual: pages with screenshots, a settings reference for every screen, and a "verify every claim in code" rule for the people who write it. The wiki mirrors the same pages for reading on GitHub.

The source lives in docs/ and runs on Astro Starlight. npm run dev in that folder gives you a local preview.

• Bugs and feature requests: GitHub issues. The About screen in the app has a Report a bug row that pre-fills one, and a bot turns a release crash trace into a readable one on your issue within a couple of minutes.
• Questions and ideas: GitHub Discussions.
• Chat: t.me/wmkeyboard on Telegram.
• Anything you would rather not or can't post publicly: arianmollik323@gmail.com. Security reports go through private vulnerability reporting.

You need a JDK 17 or newer to launch Gradle, and the Android SDK with compileSdk 36.1. Gradle provisions its own toolchain through Foojay on the first build. Android Studio's bundled JBR works well as JAVA_HOME.

./gradlew assembleFullIntlDebug   # every feature, every UI language

./gradlew assembleLiteEnDebug     # fewer tools, English UI, small

./gradlew unitTests               # the whole suite, ~7,000+ tests

./gradlew staticAnalysis          # detekt + Android Lint, 20+ minutes

The task name is the edition, then the UI-language build, then the build type. Two flags in local.properties matter: the build channel (Play versus direct download), which decides whether the ML models and LLM runtime ship inside the APK or download on demand, and a no-internet switch that strips the permission entirely.

app/                  Settings app, manifest, resources, bundled assets, most unit tests
feature/
├── ime/              WMKeyboardService and the Compose keyboard UI (the keyboard itself)
├── addons/           Addon install, reconcile, download
├── tools/            Network tool clients: AI, GIF and sticker, search, link preview
├── llm/              On-demand LiteRT-LM runtime (Play channel only)
├── translate/        On-demand ML Kit translator (Play channel only)
├── litert/           On-demand LiteRT interpreter: offline Whisper, sticker cutout
└── handwriting/      On-demand ML Kit ink recogniser
core/
├── settings/         SettingsRepository (DataStore), KeyboardSettings, modes, power saving
├── intelligence/     Grammar (Harper JNI), local LLM, handwriting, spell checker service
├── feedback/         Key sounds and haptics
├── voice/            SpeechRecognizer plus offline Whisper dictation
├── plugins/          Lua plugin sandbox
├── addons/           Addon store and repository data layer
├── content/          Clipboard, snippets, fonts, media, stickers
├── tools/            Offline tool engines: calc, units, currency, symbols, calendars
├── kdeconnect/       KDE Connect client
├── icons/            Icon packs and the SVG parser
├── theme/            ThemeSpec, palettes, rendering
├── emoji/            Catalog loader, semantic search, usage tracking
├── prediction/       Trie, SuggestionEngine, UserLexicon, dictionaries, gestures
├── input/            Composers (CJK, cluster scripts) and the input pipeline
├── keyman/           Keyman keyboard import
├── language/         Scripts, layouts, transliteration
├── common/           Utilities, direct boot, debug log, shared contracts
└── config/           Build flags and API keys
tools/dictc/          Host-side dictionary compiler, shares :core:prediction sources
native/               Rust sources for the Harper grammar bridge
docs/                 The documentation site (Astro Starlight)

Building · Testing · Architecture

The easiest first contributions are testing out the app and submitting bug reports or feature requests, not code. You can also update the data which lives in a companion repository, wasi-master/wmkeyboard-data:

• A word list for a language that has none, or a better one for a language that does.
• Emoji keywords in your language.
• A theme, a layout you miss from another keyboard, an icon pack.

For code: commits follow Conventional Commits, type(scope): summary. Run unitTests and staticAnalysis before you open a pull request, since CI runs only the first. Open an issue first for anything larger than a fix so we can agree on the shape of it.

Contributing guide

WM Keyboard stands on a lot of open work. The in-app About screen carries the full licence text for each of these, and app/src/main/assets/licenses/ holds the same files.

• Harper for offline grammar checking, bridged from Rust.
• whisper.cpp and whisper-android for offline dictation.
• Google ML Kit and LiteRT for handwriting, scanning, on-device translation and the local LLM runtime, in the Full edition.
• Tesseract and Leptonica for OCR beyond the languages ML Kit covers.
• Gemoji, KEmoji and Unicode CLDR for emoji names and keywords. Noto for fonts and animated emoji.
• Keyman for the keyboard import format, and Khipro for the Bangla layout of the same name.
• LuaJ for the plugin sandbox. Bouncy Castle for backup encryption.
• Dracula, Nord, Solarized, Catppuccin and Tokyo Night for the palettes behind five built-in themes.
• The word-frequency lists, offensive-word lists and jyutping tables listed in wordlist-sources.txt, redistributed under their own licences.

This keyboard was developed entirely by me, with the assistance of Generative AI for certain planning tasks and more tedious aspects of the development process.

For most features, I write the code myself, occasionally using AI-assisted code completion within my editor. For more complex features, I use Claude Code with Anthropic’s frontier models to assist with planning and, where appropriate, generate boilerplate code. For particularly repetitive or time-consuming tasks, I may delegate the implementation to Claude Code in its entirety.

Regardless of how AI is used, I remain involved throughout the development process. I review everything it generates, make the necessary decisions and adjustments, and test the resulting implementation whenever possible. AI is an assistant in the workflow, not a replacement for my involvement or responsibility for the code.

If AI-assisted code is a boundary you are unwilling to cross, I completely respect that position. There are some quite good keyboards developed without AI assistance. However, I believe it is increasingly difficult to disregard AI as a valuable and increasingly integral part of modern software development workflows. Whether we welcome it or not, AI-assisted development is here to stay.

MIT. See LICENSE.