A typing assistant for Windows that fixes your typos as you go, finishes your sentences, and learns how you write.
Everything runs on your own computer. No account, no cloud, no telemetry.

Download the installer · How it works · Privacy · Build · Architecture

▶ Watch the 40-second promo with sound

How it works

Fixes typos on the space bar	thhe becomes the, becuase becomes because, the instant you finish the word. A brief blue underline shows what moved. Press ← within 3 seconds to undo.
Finishes your sentences	Pause for a moment and grey ghost text completes the sentence. Tab accepts it and offers the next one. Ctrl+Tab takes one word. Esc dismisses.
Learns how you write	Retype a word you just closed and the pair becomes an autocorrect entry. Undo a fix twice and the word joins your dictionary. Rules you keep undoing stop auto-applying.
Works everywhere	Mail, chat, browsers, editors, terminals. It was built with Claude Code's prompt in mind.
Keeps your secrets	Card numbers, SSNs, IBANs, account numbers and password: style secrets never reach the model. Banking and health sites are off by default.

A tray icon opens a flyout with today's numbers and the switches. The dashboard shows acceptance over time, latency, per-app stats, what it learned, and exactly what the model saw for the last suggestion.

Install

1. Download CaseComplete-<version>.msi from the latest release and run it. It installs for your user only (no admin) and adds a Start Menu entry.
2. On first start it downloads the inference engine (llama.cpp) and the default model, Gemma 3 1B (about 1 GB), into %LOCALAPPDATA%\CaseComplete.
3. Click the tray icon. Turn on Start with Windows in Settings if you want it always there.

An NVIDIA GPU gives about 120 ms from keystroke to suggestion. On CPU, pick SmolLM2 360M in the Model menu.

Requirements: Windows 11 (Windows 10 should work), the WebView2 runtime (included with Windows 11).

Models

Models are downloaded at first run from Hugging Face and are not part of this repository. Switch from the tray menu or the dashboard.

Model	Maker	License	Notes
Gemma 3 1B (default)	Google	Gemma Terms of Use	best quality per millisecond on a GPU
SmolLM2 360M	Hugging Face	Apache-2.0	fastest, fine on CPU
Qwen2.5 1.5B	Alibaba	Apache-2.0	alternative

Spelling and grammar use Harper (Apache-2.0) in-process; no model needed, about 2 ms per check.

Privacy

Everything is local. The only network calls are the one-time downloads of llama.cpp (GitHub) and the model (Hugging Face).

• Text with a card number, SSN, IBAN, a long digit run, or a secret after a label such as password: gets no completion, no fix, and no log entry.
• privacy.toml in your profile lists apps and window-title fragments where the assistant stays off (banking, health portals, password managers by default). Edit it freely.
• The event log keeps typos and their fixes and the text of accepted suggestions. The sentence around a fix is stored only if you set store_context = true.
• Profile data lives in %LOCALAPPDATA%\CaseComplete\profiles\default. Delete the folder and it forgets everything.

Build from source

Requires Rust 1.93 or newer. The installer needs the WiX 5 dotnet tool: dotnet tool install --global wix --version 5.0.2.

cargo build --release
powershell -File installer\build.ps1     # builds installer\out\CaseComplete-<version>.msi

CC_CONSOLE=1 opens a console with the log; otherwise the log is %LOCALAPPDATA%\CaseComplete\casecomplete.log. Other knobs: CC_MODEL, CC_MIN_CONFIDENCE, CC_AUTO_APPLY_UP_TO (off|S0|S1|S2), CC_FIX_INJECT (keys|slow|paste|fast).

Layout

Crate	Role
cc-hook	low-level keyboard hook, shadow buffer, key capture during injection
cc-context	UI Automation: caret position, text before the caret, focused app
cc-engine	llama-server sidecar, model download, streaming completions
cc-correct	Harper-based checker with severity levels, autocorrect list, dictionary
cc-overlay	ghost-text overlay window (Direct2D)
cc-inject	SendInput and clipboard injection strategies
cc-store	events.db: every fix and suggestion with its outcome; tuning
cc-ui	local web UI (flyout and dashboard) and JSON API
cc-window	WebView2 windows for the UI
cc-tray	tray icon and menu
cc-app	the orchestrator
tools/bench	next-word accuracy and latency across models
promo	the Remotion promo video

Design notes: research and architecture, plan, brand.

Roadmap

• Per-app overrides (insertion strategy, context source)
• Tier-2 sentence rewrite on a pause, suggest-only
• Fine-tuning the completion model on your own writing (LoRA), locally
• A bench on your own text to pick the model that knows your voice

Support the project

CaseComplete is free and will stay free. If it saves you time every day, you can sponsor the work on GitHub; it pays for the hardware the models are tested on and keeps the releases coming.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md. Security reports: SECURITY.md.

License

MIT. See LICENSE.