LocalTypist

LocalTypist is a private native macOS autocomplete app inspired by Cotypist's workflow. It is built from scratch in Swift/SwiftUI and uses your local GGUF model through llama.cpp.

What Is Included

• SwiftPM macOS app project.
• Menu bar app with Settings.
• Safe auto-start of text monitoring when completions are enabled and Accessibility permission is already granted.
• Direct Gemma model auto-loads in the background by default when direct mode is selected.
• Accessibility permission onboarding.
• Setup preflight checks for Accessibility, optional Screen Recording, model readability, llama.cpp runtime libraries, and selected backend readiness.
• Model backend self-test from Settings, with elapsed time and output preview for the selected direct or server backend.
• Observer-first focused text monitoring through macOS Accessibility, with fallback polling.
• Local suggestion overlay using a non-activating NSPanel.
• Optional floating accessory button beside focused text fields for force-completing manually.
• Tested overlay layout for inline and mirror suggestions, including edge clamping and below-screen flipping.
• Local spelling-fix suggestions using macOS spell checking, with exact source-text replacement on accept.
• Local emoji shortcode suggestions such as :rocket -> 🚀, with exact shortcode replacement on accept.
• Global event tap for:
  • configurable next-word acceptance
  • configurable full-completion acceptance
  • configurable dismiss, force-complete, pause-current-app, and global toggle shortcuts
• Clipboard context toggle.
• Optional ScreenCaptureKit + Vision OCR context toggle; screen text is processed locally and not persisted.
• Obvious secret patterns are redacted from prompt context before local model generation.
• Custom AI instructions.
• Autocorrect controls for suppressing completions while typing a likely typo and showing suggested fixes.
• Mid-line completion toggle for suggesting text when there is existing text after the cursor.
• Emoji settings pane with a local shortcode suggestion toggle.
• Personalization settings pane with custom AI instructions, opt-in encrypted local typing history, and a word-choice strength slider.
• Per-app rules and temporary pause support.
• Per-app privacy controls for clipboard context, screen context, and personalization data.
• Per-app custom AI instructions that merge with global writing preferences for the current prompt.
• Menu bar quick controls for current-app enablement, mirror/pill mode, troubleshooting insertion, pause/resume, per-app privacy controls, global context toggles, and model loading.
• Compatibility defaults for terminals, browser/editor edge cases, and Electron-style apps.
• Launch-at-login settings using SMAppService.
• Local JSON completion statistics in Application Support, including generated/accepted counts and average generation latency.
• Local encrypted typing-history records in Application Support when personalization is enabled; the AES key is stored in Keychain.
• Typing-history personalization also redacts obvious secrets before encrypted local storage.
• Direct in-process libllama integration through a small C shim target.
• Vendored llama.cpp bootstrap script pinned to a Gemma 4-capable official tag; SwiftPM prefers this runtime when built.
• Cancellation-aware direct generation: queued stale requests are skipped, in-progress token generation checks cancellation, and only the newest request can publish a suggestion.
• Optional llama-server fallback backend.
• Shortcut recorder controls in the Completion settings pane.
• Build script discovers vendored or Homebrew libllama/ggml dylib dependencies, stages them into dist/LocalTypist.app/Contents/Frameworks, rewrites load paths to @rpath, and fails if Homebrew load paths remain.
• Staged app bundle includes a ScreenCaptureKit purpose string for optional screen-context OCR.
• llama-server process management and /completion API client.
• Build/run script that stages a real .app bundle.

Requirements

• macOS 14 or newer.
• Apple Silicon Mac.
• Swift 6 toolchain / Xcode command line tools.
• Homebrew llama.cpp tools, already found on this machine at:
  • /opt/homebrew/bin/llama-server
  • /opt/homebrew/bin/llama-cli
• For direct Gemma 4 E2B support, run the vendored bootstrap once:
  • ./script/bootstrap_llama_cpp.sh
• A local GGUF model file. Choose its path in Settings after first launch.

Run

cd LocalTypist
./script/bootstrap_llama_cpp.sh
./script/build_and_run.sh

Open Settings from the menu bar item, grant Accessibility permission, then start the model server. Direct libllama mode is the default; the server backend remains available from Settings as a fallback.

Notes

The first direct model load can take a while because the 3.2 GB GGUF model must be loaded into the app process. The Model settings pane lets you choose between the direct libllama backend and a llama-server fallback backend.

The installed Homebrew llama.cpp 8610 does not understand the model's gemma4 GGUF architecture on this machine. The project therefore vendors a newer official llama.cpp tag under .vendor/llama.cpp for direct mode.

This app does not copy Cotypist's proprietary executable code, bundled libraries, assets, fonts, local database, or private implementation.

Current Limitations

• The app bundle carries the direct libllama/ggml dylibs copied from the vendored llama.cpp build when present, falling back to Homebrew only if the vendored build has not been created.
• Screen OCR context depends on macOS Screen Recording permission and app/window capture compatibility.
• App compatibility will need iterative testing in the apps you actually type in.