Windows & Linux input method · 2026
Predict
Predict is a native input method — a TSF text service on Windows, an fcitx5 engine on Linux — that queries a local n-gram daemon to complete the current word, suggest the next one, autocorrect and agree while you type. No telemetry: the model, learned words and settings never leave the machine. A Windows installer ships the model and a native administration panel.
0
Network request
< 1 ms
Completion
2
OSes, 2 languages
Highlights
- 01TSF text service loaded into every Windows application: pre-edit, a Direct2D candidate bar that follows the accent color, theme and DPI, keyboard layout declared per profile.
- 02Local n-gram daemon (AF_UNIX, one JSON line): sub-millisecond completion, next word, accent restoration, gender/number agreement through the Lefff lexicon, learned words re-ranked on the model's own scale.
- 03One portable core shared by the fcitx5 engine and the TSF service, covered by tests on Linux (headless harness) and Windows (CTest, GitHub Actions CI on four distros + Windows).
- 04Self-contained Inno Setup installer (model included, AppContainer rights for Store apps) and a Win32 administration panel: language, settings, reformulation API key, daemon status.
- 05On-demand reformulation (Ctrl+Alt+R) through an OpenAI-compatible API, the only online feature — prediction stays local.
Technologies
- C++17
- Windows TSF
- Direct2D
- fcitx5
- n-gram (Kneser-Ney)
- Inno Setup
- Next.js
Gallery
Next project
WASIA SUP'