Typer

Local, on-device autocomplete for macOS. Typer shows a faint inline suggestion right at your cursor in (almost) any text field — keep typing to follow it, press Tab to take a word, or ` to take the whole thing. Everything runs on your Mac with llama.cpp and a small GGUF model — no cloud, no account, no Apple Developer Program.

🌐 typr.frgmt.xyz · 📓 Changelog · 🛠 Contributing

Status: alpha. Feels good in native and Electron/WebKit apps; terminals and custom-drawn editors still have approximate caret placement. Feedback and PRs welcome.

What it does

• Inline ghost-text completions at your caret, streaming word-by-word (first word in well under ~100ms).
• Type-into-the-suggestion — as long as you type what it predicted the ghost just shrinks; it only re-thinks when you diverge.
• Learns your voice locally — it primes the model with how you actually write. Nothing leaves your Mac; clear it any time.
• Per-app context, plus opt-in window-text, clipboard, and topic memory (all on-device).
• Typo correction via the macOS spell checker (off by default).

Key	Action
Tab	Accept the next word
` (backtick)	Accept the whole suggestion
Esc	Dismiss
(keep typing)	Follow along — matching keystrokes consume the ghost

Requirements

• macOS 14 (Sonoma) or later — Apple Silicon recommended
• Xcode Command Line Tools — xcode-select --install
• Homebrew

Install

git clone https://github.com/frgmt0/typer.git
cd typer
./install.sh

install.sh installs llama.cpp, downloads a model, writes a default config, and builds ~/Applications/Typer.app. Then launch it:

open ~/Applications/Typer.app

On first launch, onboarding walks you through permissions and picking a model size.

Permissions (System Settings → Privacy & Security)

• Accessibility — required. How Typer reads what you type and inserts suggestions.
• Screen Recording — optional. Only for caret placement in terminals + on-screen context.

Onboarding links you straight to the right pane. After a rebuild you may need to re-grant Accessibility — see stable signing to make the grant stick.

Using it

Click the ⌨︎ menu-bar icon for everything: toggle features live, switch model size, see accept-rate stats, clear your learned style, open the config/log, or check for updates.

• Model size — pick Small (fast, runs on any Mac) or Large (higher-quality suggestions, best on 16 GB+ of RAM). Large downloads once, on demand, and switches live — no restart. Change it from the menu dropdown or during onboarding.

Updating

Typer builds from source, so updating means pulling the latest and rebuilding. Either:

./update.sh        # fast-forward to the latest, rebuild, relaunch

…or click the ↻ button in the menu-bar popover — it fetches, tells you how many commits behind you are, then rebuilds and restarts in the background (progress in ~/Library/Logs/Typer-update.log). It only fast-forwards, so local changes are never overwritten.

Privacy

Everything is local — the model runs on your machine and nothing is sent anywhere.

• Secure input is never captured — password fields, the login window, sudo, password managers: no buffering, learning, logging, or generation.
• The log is not a keylogger — by default it records no typed text (only counts/events).
• Your files are yours — style.txt and stats.json live under ~/Library/Application Support/typer/ (mode 0600); wipe them any time (Menu → Clear Learned Style / Reset All Data).

Get your AI coding agent up to speed

Hacking on Typer with Claude Code, Cursor, or similar? Paste this to orient your agent fast:

You're contributing to Typer, a local on-device autocomplete app for macOS (Swift/AppKit + SwiftUI front end, a C++ llama.cpp helper, and a Python training pipeline). Get oriented before we change anything:

1. Read README.md and CONTRIBUTING.md.
2. Map the architecture: the menu-bar app in scripts/typer/ — start at TyperApp.swift (the core TyperApp class) and its extension TyperApp files split by concern (+EventTap, +Completion, +Caret, +Context, +Menu, +Model); the generation helper scripts/llama_server.cpp (llama.cpp over a JSONL stdin/stdout protocol); and the model work in training/ (read training/README.md).
3. Note the flow: scripts/build.sh compiles the Swift app + C++ server, signs, and installs to ~/Applications/Typer.app; ./update.sh pulls + rebuilds.

Then give me a short summary of (a) how a keystroke becomes ghost-text at the caret, (b) how the model is selected and loaded (Small vs Large), and (c) where completion quality is tuned. Ask before editing anything.

Contributing

PRs welcome — see CONTRIBUTING.md for the build/run flow, project layout, architecture, configuration, and the model-training pipeline.

License

MIT — see LICENSE. Bring your own model (subject to that model's license).