Aegis IME

Aegis is an offline-first Android input method for Simplified Chinese and English. It is built on the open rime-wanxiang CC BY dictionaries with a self-built decoder; no rime / librime at runtime. Everything you type stays on your device: no network request is ever made while you type, and nothing you type is ever sent.

English | 简体中文

Contents

• For users: install & enable
• Features
• Privacy & permissions
• Build (for developers)
• Release dictionary pack
• Architecture
• Licensing & acknowledgments

For users: install & enable

Aegis is not on an app store yet; it is distributed as a downloadable APK.

Requirements

• Android 14 or newer (minSdk 34). Older Android versions are not supported.
• A 64-bit ARM device (arm64-v8a); the published APK carries that ABI only.

Install

1. Open the Releases page and download the APK asset from the latest build. Releases are marked pre-release, and the APK is a release build signed with an Android debug key.
2. Because the app is installed outside an app store, Android will ask you to allow installing unknown apps for the browser or file manager you used; grant it when prompted, then open the downloaded APK to install (this is a sideload).

Enable Aegis as a keyboard

Menu names vary slightly by device, but the flow is the standard Android one:

1. Settings > System > Languages & input > On-screen keyboard (on some phones: Manage keyboards / Virtual keyboard).
2. Turn Aegis on. Android shows a warning that a keyboard can read what you type; this is the standard notice for every input method; see Privacy & permissions for exactly what Aegis does and does not do.
3. Switch to Aegis while typing: tap the input-method / globe key on the keyboard, or open the "Choose input method" notification / picker and select Aegis.

First run

English typing works straight away; Chinese needs one download first. The APK carries no Chinese dictionary, and the keyboard never fetches the dictionary pack on its own. When you type Chinese with no pack installed, the candidate strip offers the download — a ~102 MB transfer that expands to ~272 MB in the app's private storage, not restricted to Wi-Fi — and it starts only when you tap there, or on the dictionary card in the settings screen. An interrupted transfer resumes where it stopped, and the pack is checked against its SHA-256 before it is installed. Until it finishes, Chinese input stays locked, while English and every panel keep working.

The settings screen additionally offers an optional enhancement model (the ~420 MB octagram grammar) for sharper next-word and whole-sentence ranking; that one is fetched only when you tap to start it. Once a file is installed, its card offers a check-for-updates button, and a check that finds a newer file continues straight into the download. Beyond those two downloads and the update checks that go with them (a small metadata file for the dictionary, a HEAD request for the model), Aegis's own code makes no network requests, and nothing Aegis does reaches the network unless you ask it to.

Features

• Keyboards: 26-key full pinyin, 9-key T9 and an English 26-key layout, plus number, symbol and numeric-keypad layouts, on a self-drawn View+Canvas keyboard. In landscape the keyboard keeps its portrait width and docks to one side, so the app behind it stays reachable.
• Lattice Viterbi decoder: over a memory-mapped dictionary, scored by unigram log-probability plus a character-bigram context model; with the enhancement model installed, a further next-word and whole-sentence reranking pass runs on top.
• Fuzzy pinyin: ten rules — zh/z, ch/c, sh/s, ang/an, eng/en, ing/in, n/l, f/h, l/r, k/g. The master switch is off until you turn it on, and each rule then has a switch of its own; changes take effect immediately.
• Jianpin (initials abbreviation): zg -> 中国, bjdx -> 北京大学.
• Syllable control: on the 9-key keyboard a side column lists the readings for what you typed and locks the one you pick; picking a locked reading again offers the characters that share it. While composing, a segmentation key splits a run of letters where you want it.
• CN-EN mixed input: commit English words (e.g. wifi) without switching language.
• On-device learning: frequently and recently used words rank up and next-word pairs are learned. Automatic learning has its own switch, and next-word suggestions in the candidate strip have another one that is off until you turn it on. Automatically learned entries fade out after months of disuse; words you add yourself are never forgotten. No word is learned in a password field, or in any field that asks for no personalized learning; in those fields the symbol and emoji panels do not count what you pick either. Everything lives in the app's private filesDir (userdb.txt and userlearn.txt), and the user-dictionary screen lets you search, add, delete, import, export and clear it.
• Emoji panel: a Common tab plus categories (smileys, gestures, flags, animals, plants, food, travel, activities, objects, symbols), with skin-tone and gender variants and a lock toggle for picking several in a row.
• Clipboard history: clips you copy are kept for reuse, with per-item and batch delete, select all, and clear. The history can also be switched off entirely.
• Saved phrases (常用语): reusable phrases in categories you name, reorder and move between, each with an optional note, importable and exportable as a plain text file.
• Symbol & edit panels: a symbol board with a Common tab plus Chinese, English, currency, web, math, Greek, arrow, super/subscript, numbering, IPA and pinyin categories and a lock toggle; and a text-editing panel with cursor movement, line start/end, selection, select-all, copy, cut, paste and delete.
• Custom symbols: choose which punctuation sits in the 9-key side column and which operators sit on the numeric keypad.
• Symbol & emoji associations: typing pinyin can offer matching symbols and emoji directly in the candidate strip.
• Calculator: type an expression and the candidate strip offers its result; + - * / × ÷, parentheses, signs and % are understood.
• Copy bar: text you just copied or cut is offered back above the keyboard, either whole or split into pieces you pick from.
• Encrypted backup: export your learning dictionary, auto learning, saved phrases, clipboard history, symbol history, emoji history and all settings into a single encrypted file (AES-256-GCM, key derived with PBKDF2-HMAC-SHA256 at 600,000 iterations), and import it back by overwriting or merging. A default backup password can be kept on the device behind biometric or screen-lock authentication. The downloadable dictionary and model are not in the backup; they can be downloaded again.
• Simplified-Chinese normalization: every candidate Aegis proposes is Simplified. Traditional and variant character forms from the upstream data are folded to their Simplified image (with frequency merging) when the dictionary pack is built, using the OpenCC tables in tools/t2s-data. That happens on the build host; no conversion table ships in the APK.
• Material 3 settings screens: input settings, dictionaries & downloads, user dictionary, data backup, about & enable, and an open-source licenses page.

Privacy & permissions

A keyboard sees everything you type, so trust is the whole point. Aegis is built so that trust does not depend on our word alone:

• The app's own manifest declares two Android permissions: INTERNET and USE_BIOMETRIC. The installed APK lists a third, com.aegis.ime.DYNAMIC_RECEIVER_NOT_EXPORTED_PERMISSION, which the AndroidX libraries add when the manifests are merged; it is not a platform permission, is declared at signature level in Aegis's own namespace, and asks for nothing on your device.
• INTERNET fetches the dictionary pack (offered by the keyboard while no pack is installed, because no Chinese dictionary ships in the APK, and started only when you tap), the optional enhancement model (only when you tap to start it), and the update checks for those two. No network request is ever made while you type, and nothing you type is ever sent.
• USE_BIOMETRIC is used only for the default backup password: saving it, or filling it into a backup dialog, needs a biometric or screen-lock confirmation first.
• Your keystrokes, candidates, learned words, user dictionary, and clipboard never leave the device: they live in the app's private storage (filesDir). The only way any of it leaves is a file you export — a backup, your learning dictionary, or your phrases — to a location you pick.
• Aegis is excluded from Android's cloud backup and device-to-device transfer, so its data is not carried off the device that way either.
• There is no analytics, no telemetry, and no account.
• One part of the APK is not Aegis's own code: androidx.emoji2 arrives as part of Jetpack Compose and, once one of Aegis's own screens has been opened, looks for a package in the system image that offers an emoji font and asks it for one; where the device has no such package, nothing is asked. It does this at most once per run of the app, and typing does not trigger it. Aegis opens no connection for it and sends nothing of yours with it.

See PRIVACY.md for the full statement.

Build (for developers)

Prerequisites:

• JDK 25 — the toolchain CI builds with; Java and Kotlin bytecode target 17
• Android SDK with platform 37 and build-tools 36 (compileSdk / targetSdk are 37; minSdk is 34)
• Gradle is provided by the wrapper; just use ./gradlew (no system Gradle needed)

./gradlew assembleDebug           # build the debug APK
./gradlew :app:testDebugUnitTest  # JVM unit tests (decoder, dict, fuzzy, jianpin, learning, UI)
./gradlew :app:lintDebug          # Android lint

No dictionary-derived asset is packaged in the APK at all: app/build.gradle.kts excludes aegis_dict.bin, aegis_t9.bin, aegis_jianpin.bin and the character-bigram context model aegis_lm.bin from the packaged assets. All four are downloaded at runtime into filesDir/downloaded/ as members of one dictionary pack; the three dictionaries are the only source of Chinese candidates and aegis_lm.bin reranks them. Decoding does not strictly require aegis_lm.bin, but a pack missing it counts as incomplete and the app offers the download again.

The decode tests that need the real tables read them from the directory named by AEGIS_FULLDICT_DIR; python3 tools/fetch_test_dict.py downloads the published pack and unpacks it into app/src/main/assets/, where it stays out of both the APK and git.

The pack is prebuilt by the :tools module from all 14 wanxiang tables (zi jichu lianxiang cuoyin duoyin shici diming yixue huaxue yaopin mingren yiren wuzhong renming) at --min-freq 1 with no per-key cap, so it keeps every entry.

./gradlew :tools:installDist
tools/build/install/tools/bin/tools --out <dict> --min-freq 1 --keytype letter   --t2s-data tools/t2s-data <14 wanxiang .dict.yaml ...>
tools/build/install/tools/bin/tools --out <t9>   --min-freq 1 --keytype digit    --t2s-data tools/t2s-data <14 ...>
tools/build/install/tools/bin/tools --out <jp>   --min-freq 1 --keytype initials --t2s-data tools/t2s-data <14 ...>
tools/build/install/tools/bin/tools lm --out <lm> --t2s-data tools/t2s-data <14 wanxiang .dict.yaml ...>

The screenshots in this README are rendered, not photographed: the unit tests write them to app/build/render/i18n/{en,zh}/, and docs/screenshots/ holds the copies linked above.

Release dictionary pack

App APK releases and downloadable dictionary packs are published separately. Versioned app releases carry the APK only. The dictionary pack is published on the rolling dict-latest GitHub release, and the app discovers dictionary updates from that single release tag by comparing the installed pack's SHA-256 with the current dictionary ZIP asset, or by noticing that the installed pack is missing one of its members.

tools/release/build_dictionary_pack.py --release-tag dict-latest

The command clones amzxyz/rime-wanxiang — the wanxiang branch by default, --source-tag to pin a release tag instead, --source-dir to use an existing checkout — builds the 14 verified tables with tools/DictBuilder at --min-freq 1 and no per-key cap, and writes the pack ZIP, aegis-build-info.json, and aegis-dictionary-update.json under build/release-dictionary/. Upload those generated files to the rolling dict-latest release, not to versioned app releases. The checked-in aegis-build-info.json records the trail of the pack it was generated from (source tag & commit, per-table input hashes, build parameters, output-bin hashes, and physical asset URL) and its remaining provenance gaps; regenerate it whenever the rolling dictionary pack is republished. A dictionary pack is not a fully reproducible public supply-chain artifact until the release also carries the exact input hashes, a deterministic recipe, and a signature or attestation.

Architecture

• app/.../ime: IME service, self-drawn keyboard/candidate views, panels (emoji, clipboard, symbols, custom symbols, edit), copy bar, input state machine.
• app/.../decoder: PinyinDecoder (word-lattice Viterbi; exact > fuzzy > jianpin edges).
• app/.../dict: mmap readers BinaryDict, CharBigramLM, Fuzzy, OctagramReader, and ModelDownload — the one place any network request is made.
• app/.../engine: candidate assembly, symbol/emoji associations, calculator.
• app/.../layout: key and layout definitions, symbol and emoji catalogs.
• app/.../user: on-device data — UserModel and UserLearning, clipboard and phrase store, custom symbols, symbol usage.
• app/.../backup: encrypted backup archive, crypto, restore.
• app/.../ui: Compose settings screens, dictionary/model cards, user dictionary, backup, about, licenses.
• tools/: host dict/LM builders (DictBuilder, LmBuilder) + eval verifier + release packager.

Licensing & acknowledgments

Aegis's own code is GPL-3.0 (see LICENSE). Aegis ships a self-built decoder (clean-room Kotlin) and is an independent project, not affiliated with the RIME project; it links no librime / native code. It stands on open data from the wanxiang project by amzxyz, with our deepest thanks. The third-party attribution / ShareAlike obligations below are not waived by Aegis's own license.

rime-wanxiang dictionaries: (c) amzxyz and rime-wanxiang contributors, CC BY 4.0 (rime-wanxiang). The downloadable aegis_{dict,t9,jianpin}.bin and aegis_lm.bin are derivatives of the full 14 tables (字 基础 联想 错音 多音 诗词 地名 医学 化学 药品 名人 异体 物种 人名). Changes: tones stripped (ü -> v), syllables concatenated into toneless keys, traditional/variant forms folded to Simplified (OpenCC tables), repacked into Aegis's binary format; the pack keeps every entry (--min-freq 1).

wanxiang octagram model (wanxiang-lts-zh-hans.gram): (c) amzxyz, CC BY 4.0 (RIME-LMDG). The optional top-tier context model behind next-word / whole-sentence ranking; fetched only on explicit opt-in, not bundled in the APK. OctagramReader (Kotlin, GPL-3.0) is original Aegis code; its on-disk format was clean-room reverse-engineered from librime-octagram (GPL-3.0) + darts-clone; no upstream source copied.

OpenCC data (tools/t2s-data): traditional-to-simplified and variant mappings, Apache-2.0 (see tools/t2s-data/LICENSE-OpenCC and tools/t2s-data/PROVENANCE.md); used at dictionary-build time only.

Other: AndroidX / Jetpack Compose / Material 3 / Kotlin stdlib: Apache-2.0; JUnit: EPL (test scope only, not distributed). Algorithm references (not vendored): AOSP PinyinIME (Apache-2.0), darts-clone (BSD-2-Clause).

The full list, with the modifications made to each work, is in THIRD_PARTY_LICENSES.md; the app shows the same list under Settings → About & enable → Open-source licenses.