Lumen

Browse. Remember. Ask. No cloud required.

Demo • How It Works • Knowledge • Privacy • Stack • Building

---

I kept reading great stuff online and then completely blanking on where I saw it, and so I built a browser that remembers for me.

Lumen reads along as you browse. It grabs the good parts of every page you actually spend time on, summarizing them, and filing them into a knowledge base that lives entirely on your phone. You can ask about any of it later. A local LLM answers from your reading, not the whole internet, and it never calls any external services.

Runs on any iPhone or iPad on iOS 18+. No account, no server, no "sign in to continue." You browse like normal and it does the rest.

Demo

     

How It Works

  You browse the web
         │
         ▼
┌─────────────────┐      reading signals detect
│   Lumen reads   │ ◀─── when you actually engage
│   along with    │      with a page, not just
│   you           │      open it
└────────┬────────┘
         │
         ▼
┌─────────────────┐      content extracted,
│  Knowledge DB   │ ◀─── embedded, summarized,
│  SQLite + FTS5  │      classified — all on-device
└────────┬────────┘
         │
    ┌────┴────┐
    ▼         ▼
 📂 Browse  ✦ Ask
 Topics →   "What did I
 Sites →     read about
 Pages       closures?"

Knowledge System

The knowledge system has two tabs:

✦  AI Chat Ask like you'd text a friend. Lumen finds the pages that matter, hands them to a local Llama 3.2 1B, and answers from what you read online instead of internet surfing. Every answer shows its sources, so you can always check its validity.

📂  Folders Everything you read sorts itself into folders: Topics → Websites → Pages Every layer writes its own little summary. Topics get sorted automatically, and each site gets a synthesis of everything you read there.

┌─────────────────────────────────────────┐
│  EVERYTHING RUNS LOCALLY                │
│                                         │
│  LLM inference    ████████  MLX Swift   │
│  Embeddings       ████████  NLEmbedding │
│  Full-text search ████████  FTS5        │
│  Vector search    ████████  Cosine sim  │
│  Storage          ████████  SQLite      │
│                                         │
│  No networking.                         │
└─────────────────────────────────────────┘

Privacy

There's no server that we send anything to. That's kind of the whole point :)

Layer	Protection
Network	HTTPS-only upgrades, mixed-content blocking
Cookies	Third-party cookies blocked by default
Tracking	Built-in tracker database with threat classification
Fingerprinting	Fingerprint resistance via content security policies
Data	All knowledge stays in local SQLite
AI	LLM runs on-device via MLX

Stack

Swift 6.2 · SwiftUI / iOS 18+ · Xcode 26.2+
│
├── 🧠  MLX Swift ──────── on-device Llama 3.2 1B inference
├── 🌐  WKWebView ──────── hardened browser engine
├── 💾  SQLite + FTS5 ──── full-text search & content tables
├── 🔢  NLEmbedding ────── Apple's sentence-level embeddings
├── 🛡️  ThreatDetector ─── tracker & fingerprint classification
│
└── Zero external dependencies beyond Apple + MLX

Building

Requirements

• macOS with Xcode 26.2 or newer
• iOS 18 or newer device or simulator (Apple Silicon Mac required for the simulator)
• Apple Developer account for code signing
• Network access on first launch (the LLM weights are pulled from Hugging Face)

Steps

# clone
git clone https://github.com/Lux-Softworks/Lumen.git
cd Lumen

# open in Xcode
open Lumen.xcodeproj

In Xcode:

1. Select the Lumen target → Signing & Capabilities.
2. Replace the bundled team (XF6K537DNY) with your own, and change the bundle identifier from com.luxsoftworks.Lumen to something unique to you (e.g. com.yourname.Lumen). Do the same for the LumenTests and LumenUITests targets.
3. Swift Package Manager will resolve the MLX Swift dependencies automatically on first open.
4. Pick a destination (iOS 18+ device or iOS 18+ simulator on Apple Silicon) and hit ⌘R.

First launch

The first time you open the knowledge panel, Lumen downloads the mlx-community/Llama-3.2-1B-Instruct-4bit weights (~700 MB) from Hugging Face and caches them on-device. After that, everything runs fully offline.

Deployment

Lumen is a client-only iOS app. "Deploying" means getting the build onto a device:

• Run on your own device — connect an iPhone/iPad (iOS 18+), select it as the destination, and ⌘R. Trust the developer profile under Settings → General → VPN & Device Management on first run.
• Share via TestFlight — in Xcode, Product → Archive, then distribute the archive to App Store Connect and invite testers through TestFlight.
• App Store release — submit the same archive for App Store review. Distribution through Apple's App Store is explicitly permitted by the license exception below.

License

AGPL-3.0 with an Apple App Store distribution exception — if you build on this, share it back.

The exception (added as additional permission under GNU AGPL version 3 section 7) authorizes distribution of this software through Apple's App Store under Apple's terms. All other distribution remains governed by the AGPL-3.0.

Contributing

Want to help? Awesome — here's the process:

1. Fork the repository and create a feature branch off main (git checkout -b your-feature).
2. Match the conventions — read CLAUDE.md for the project's code style. Most importantly: this codebase contains self-explanatory code (meaning no comments), and all building/testing happens in Xcode (build with ⌘R, run tests with ⌘U).
3. Test on a physical device for anything AI-related. The on-device LLM does not run in the Simulator.
4. Open a pull request against main with a clear description of what changed and why.

Thanks for helping improve our community and software!

AI Declaration

Lumen was built entirely by me as a solo developer. I used AI tools (primarily Claude) throughout development for brainstorming architecture decisions, debugging, generating boilerplate, and writing code. All design decisions, system architecture, and feature direction are my own. I reviewed and understand every line of code in the project.