Lumen vs Magic AI Keyboard

Lumen is a browser with a memory. Most browsers forget the second you close a tab. Lumen does not. As you read, it quietly builds a personal knowledge base from the pages you actually engage with, and then lets you ask questions about any of it in plain English, answered by an AI model that runs entirely on your device. Nothing you read, nothing you ask, and nothing Lumen learns about you ever leaves your phone. WHAT IT DOES Lumen reads along with you. It distinguishes between pages you genuinely read and pages you opened and swiped away, and it extracts the meaningful content from the ones that stuck. That content gets summarized, embedded, and organized into a searchable knowledge base, all on-device. Later, when you want to find something, you do not have to dig through history or remember which tab it was in. You just ask: "What did I read about sourdough starters last week?" "Summarize that article about the new M-series chips." "What were the main arguments in that piece I read on Tuesday?" Every answer cites the pages it came from, so you can jump straight back to the source. AUTO-ORGANIZED Your reading sorts itself into a clean hierarchy: Topics, Websites, and Pages. Each level carries its own AI-generated summary, so you can scan what you have been learning at a glance. No tagging, no folders to maintain. Lumen handles it in the background. PRIVATE BY DEFAULT Lumen has no server. There is quite literally nothing to send. On-device LLM: a quantized Llama 3.2 model runs locally through Apple's MLX framework. Local SQLite storage: your reading history, summaries, and embeddings never leave the device. Built-in privacy: tracker blocking, third-party cookie blocking, and fingerprint resistance. Secure: HTTPS-only by default, with mixed-content blocking. No strings attached: no accounts, no sync, no telemetry, no ads. The AI weights download once on first launch (~700 MB). After that, everything works fully offline: browsing, reading, asking, answering. BUILT FOR iOS Lumen is written from scratch in Swift and SwiftUI, designed specifically for iPhone and iPad. It uses Apple's own frameworks wherever it can: NLEmbedding for semantic search, WKWebView for browsing, SQLite with FTS5 for full-text queries, and MLX Swift for on-device inference. Zero third-party trackers. Zero external services. WHO IT'S FOR While most browsers treat the web as a feed to scroll past, Lumen treats it as a library worth keeping. If you read a lot online and constantly think "I know I saw something about this recently but I cannot find it again," Lumen is for you. Researchers, students, writers, developers, and lifelong learners. You do not need to change how you browse. Browse normally, and Lumen takes care of the rest. REQUIREMENTS iOS 18 or later. iPhone with Apple Silicon recommended for the fastest on-device AI performance.

Turn rough thoughts into clear writing without leaving the app you're in. Magic AI Keyboard is an iOS keyboard built around one simple action: Magic. Type what you mean — even if it's rough, awkward, or unfinished — then tap Magic. It gives you a cleaner, more natural version while preserving your meaning. No copy-and-paste trip to another AI app. No complicated prompt. Just type, tap, and keep going. A keyboard first • English QWERTY • Japanese flick input • Japanese romaji input • Normal typing works without AI • On-device Japanese conversion Magic when you want it Magic can clean up grammar, awkward phrasing, and clarity without turning writing into a chat with an AI assistant. You stay in control Magic never sends a message for you. AI processing happens only when you explicitly use Magic. Keyboard input is not continuously collected or sent. Some AI and cloud conversion features require Full Access. Normal typing and on-device Japanese conversion remain available without it. Write first. Tap Magic when you need it. You decide when to send. We utilize Google and OpenAI APIs to provide predictive text suggestions and AI-driven analysis terms: https://yama-code.dev/terms privacy: https://yama-code.dev/privacy