Lumen

Local-AI autocomplete + spell-correction for macOS — powered by MLX. Runs 100% on your Mac (Apple Silicon, Apple's MLX + Metal), lives in the menu bar, accept with Tab.

Sister project of Sombra (which uses llama.cpp). Lumen is MLX-only — an Apple-native engine, a clean macOS sidebar UI, and models chosen straight from the Hugging Face mlx-community catalog.

---

🔒 Privacy — it stays on your machine

• Inference is 100% local (MLX + Metal). No cloud, no telemetry.
• Lumen only touches the internet on your action: downloading an MLX model from Hugging Face. Everything else runs offline.
• The code is open — audit it. You can block all connections with a firewall (e.g. LuLu) and autocomplete keeps working; you just won't be able to download new models.

Features

• Word-by-word autocomplete — each Tab inserts the next word; a sticky multi-word suggestion lets you Tab through a whole phrase.
• Instant word completion from an embedded dictionary (no GPU) while the model predicts the next phrase at word boundaries.
• Spell correction of the last word (native macOS spell-checker) — the typo is shown struck through with the fix next to it; colors are configurable.
• MLX model catalog — pick a recommended model (Qwen2.5 0.5B, Qwen3 1.7B/4B, Gemma 3 1B…) as a card, or paste any mlx-community/… id. Downloads on first use to ~/Library/Application Support/Lumen/models.
• Per-model AI instructions — tell the AI your language, tone and role; saved per model, used as context for each suggestion.
• Learn from your typing (optional, local) — favors the words you actually use.
• Emoji shortcuts — type :name (:fire, :rocket, :festa…) with configurable gender and skin tone.
• Per-app blocking — turn Lumen off in specific apps.
• Configurable — suggestion length, temperature, context window, accept shortcuts (Tab + accept-whole), bubble color/opacity/position.
• Clean native UI — macOS-style sidebar, light & dark, amber accent.
• Bilingual UI (English / Portuguese) following your system language.

Requirements

• macOS 14+ and Apple Silicon (M1–M4)
• To build: Xcode + the Metal Toolchain component (xcodebuild -downloadComponent MetalToolchain)

Build & run

./scripts/bundle.sh release    # build & package Lumen.app (MLX compiled in)
open ./build/Lumen.app          # pick a model on first launch

Lumen ships without a model — pick one from the catalog in Preferences → Model and it downloads from Hugging Face on first use.

Grant Accessibility (System Settings → Privacy & Security → Accessibility → Lumen). That is the only permission required.

Note on MLX: the app embeds Apple's MLX engine. The Metal kernels are shipped as a prebuilt mlx.metallib (in Frameworks/, copied next to the binary by bundle.sh). Rebuilding it requires the Metal Toolchain.

Architecture (Sources/Lumen)

Component	Role
MLXPredictor.swift	inference over mlx-swift — predict (generation) + rank (scores dictionary candidates by context)
SuggestionEngine.swift	event loop: read → predict → show bubble → accept on Tab
MLXCatalog.swift	curated mlx-community model list
AXReader / KeyTap / TextInjector	accessibility read, Tab capture, text injection
GhostOverlay / IndicatorOverlay	suggestion bubble + presence badge
SpellCorrector / WordDictionary / WritingProfile	correction, instant completion, personalization
SettingsView.swift	native sidebar Preferences

License

MIT © 2026 André Luís Alves Campos. Inference uses MLX and mlx-swift (MIT). Models have their own licenses and are downloaded by the user.