ScrollCollector 📜🤖

ScrollCollector is a privacy-first, zero-latency Chromium browser extension that extracts, redacts, and classifies text content in real time as you scroll through social feeds and web pages. By default it runs entirely on your machine: a small sentence-embedding model bundled with the extension (System 1) decides every post, the browser's built-in model (Gemini Nano) is asked only about the posts System 1 is unsure of, and a keyword engine fills any gap. Cloud decision APIs are off unless you turn one on.

It is designed as a high-performance keyboard-driven curator to capture, filter, and compile web insights into interactive offline feed diaries.

---

Key Features

• System 1, bundled local embeddings: every post is first scored by a small q8 sentence-embedding model (all-MiniLM-L6-v2 by default, bge-small-en-v1.5 selectable) run with Transformers.js on WebAssembly in the extension's offscreen document. The weights ship inside the extension; nothing is downloaded at runtime.
• On-device LLM only for uncertain posts, no model downloads: posts System 1 answers with low confidence go to the browser's built-in Prompt API (Chrome Gemini Nano / Edge Phi-4-mini). The Rule-based Classifier (always available offline) fills whatever is still unanswered. The extension never downloads or bundles a generative model of its own.
• Cloud decisions are opt-in: TypeSafe Jev and OpenAI are off by default. Each needs its own opt-in and your own API key; only posts System 1 is unsure of are sent, to one primary provider at a time, never chained.
• Dynamic Tag Generation: Automatically extracts and suggests new category tags (e.g. Cooking, Space, Tesla) for items that do not match existing tags.
• Priority Routing: Custom and default tags always take classification priority over dynamic tags.
• Segmented Workspace: A clean, modern tabbed interface (Stream, Export, Settings) replacing vertical collapsible widgets to optimize narrow side panels.
• List Virtualization: Built-in IntersectionObserver-based DOM windowing. Off-screen cards are cleared from the DOM and replaced with height-matched spacers to maintain smooth scrolling and ultra-low memory usage even with thousands of items.
• Double Filtering System:
  • Category Tags Bar: Horizontal scrollable pills showing category count badges. Enabled categories are always visible even with 0 counts.
  • Websites Bar: A second horizontal scrollable pills row to filter items by source domain (LinkedIn, X/Twitter, YouTube, Medium, Facebook, Instagram, or custom sites).
  • Supports click-and-drag horizontal mouse-scrolling on both scrollers.
• Active Tab Auto-Detection: Auto-detects when you change active browser tabs and automatically selects the corresponding website filter pill (e.g. switching to a YouTube tab focuses the YouTube feed, switching to LinkedIn focuses LinkedIn items), falling back to "All Sites" on untracked domains.
• Instant Activation (No Refresh Required): Programmatically injects content scripts into existing open browser tabs matching both root domains and subdomains immediately upon extension installation or reload.
• Scroll-Optimized Observers: Uses globally-cached states inside content scripts to evaluate pause states and site activation synchronously in memory, ensuring scrolling remains lag-free and free from storage API throttling.
• Obfuscation-Proof Selectors: Uses modern HTML React data attributes (like [data-urn*="urn:li:"] for LinkedIn) to detect posts, remaining completely immune to dynamic class obfuscations.
• Native Favicon Collector: Automatically resolves and displays the site's official favicon in each card header using Chrome's offline _favicon API.
• Keyboard-driven Curation (HUD):
  • ↑ / ↓ or Tab / Shift+Tab to focus cards (scrolls into view smoothly).
  • ← / → to cycle the category tag of the focused card in real time.
  • Space to toggle Favorite (<3) status.
  • Enter to expand or collapse card snippet body.
  • Shift+Enter or Ctrl+Enter to open card link in a new tab.
• Favorites (<3): Mark items as favorites (❤️), which records a custom favoritedAt timestamp and preserves items during "Clear stream" triggers.
• Master Pause Switch: Toggle button in the header that pauses scanning, showing a banner: "Collection paused since HH:MM:SS".
• Ignored Keywords: Settings form to add keyword filters. Incoming posts containing ignored words are skipped.
• Interactive HTML Diary Export: One-click action to download your collected feed as a standalone, styled HTML file with a responsive card layout, theme toggles, search filtering, and category controls.
• Privacy First: Built-in inline PII sanitizer scrubs emails, credit cards (Luhn-validated), and phone numbers before text is processed.
• Subtle Gradient Border Glow: News-type cards automatically receive a premium multi-color gradient border glow that respects the card's rounded borders.
• No Overflow Cards: Modern CSS styling ensures long continuous strings wrap gracefully and platform links display with clean ellipsis.

---

How to Install (Local unpacked)

ScrollCollector runs locally by default (post text leaves your device only if you turn on a cloud provider). To load it into any Chromium-based browser (Chrome, Edge, Brave, Arc, Opera):

Standalone Installer (1-Click Setup)

ScrollCollector features a unified standalone setup installer built in Go.

1. Windows: Double-click ScrollCollectorInstaller.exe in the root folder.
2. macOS: Compile and run ScrollCollectorInstaller (see building instructions below).
3. The installer runs a temporary local server and opens your browser to an interactive setup wizard:
   • It auto-detects browser installations (Chrome, Edge, Brave).
   • If multiple browsers are found, it presents a checkbox list to choose your targets.
   • It extracts extension files to %USERPROFILE%\ScrollCollector (Windows) or /Applications/ScrollCollector (macOS) and copies the installation path to your clipboard.
4. Follow the interactive guide in your browser to load the unpacked extension.

Storage (no companion app)

All collected items live in the browser's own chrome.storage.local. There is no background service, no sync server, and no cross-browser synchronization — each browser profile keeps its own independent stream.

Note: earlier revisions shipped a system-tray companion app that provided a desktop dashboard and cross-browser sync. It was removed in commit 3416521 ("refactor: remove companion app") and the extension contains no runtime references to it. The Go sources remain under companion/ unbuilt, in case the feature is revived; they are not part of the extension.

To keep the stream bounded, the collector caps stored items (see MAX_STACK_ITEMS in background.js) and evicts the oldest entries first. Favourited items (<3) are never evicted.

Building from Source (Developers)

Run the compilation script in your shell to build the standalone installers:

./build-installer.sh

Note: macOS systray compilation requires CGO/Xcode clang compiler.

Manual Installation (All Platforms)

1. Download/Clone the Repository:

   git clone git@github.com:fauzulkc/ScrollCollector.git

2. Load into Browser:

   • Open your browser and navigate to chrome://extensions/ (or edge://extensions/ / brave://extensions/).
   • Enable Developer Mode (toggle in the top-right corner).
   • Click the Load unpacked button in the top-left.
   • Select the built dist/ext/ folder (run npm run fetch:models, then npm run build:ext; see BUILD.md). The source folder has no bundled model weights, so System 1 cannot start from it. On Windows with the repo in WSL, run npm run sync:win and load D:\code\ScrollCollector-dev instead: Chrome is unreliable loading from \\wsl.localhost. Or unzip releases/ScrollCollector-<version>-store.zip from npm run package:store.

3. Open the App:

   • Click the extension puzzle icon in the toolbar.
   • Click ScrollCollector to open the side panel.

---

Enabling On-Device AI

ScrollCollector decides each post with a cascade (lib/decision/cascade.js): a cache, a noise prefilter, then:

Step	Engine	When it runs	Browsers
System 1	Bundled q8 sentence embeddings (MiniLM default, bge-small selectable), offscreen document	Every post, in one batch	All Chromium browsers
Escalation	Browser Prompt API (built-in model), the default; or TypeSafe Jev / OpenAI if you opt in	Only posts System 1 answered with low confidence; one primary provider, within a rate budget	Chrome (Gemini Nano), Edge (Phi-4-mini)
Rules	Rule-based Classifier	Fills anything still unanswered	All browsers (keyword fallback)

The flags below are needed only for the escalation step; System 1 works without them.

The header status pill shows which tier is active: green = semantic AI ready, orange = downloading or keyword fallback.

Chrome 138+

ScrollCollector uses the LanguageModel global (Prompt API). No origin-trial permissions are required.

1. Chrome version: Prompt API for extensions requires Chrome 138+ on desktop.

2. Enable flags (if not already on):

   chrome://flags/#prompt-api-for-gemini-nano
   

   Select Enabled (or Enabled for Extension if shown).

   chrome://flags/#optimization-guide-on-device-model
   

   Select Enabled BypassPerfRequirement.

3. Relaunch Chrome, verify at chrome://on-device-internals, then reload the extension.

4. Header should show Gemini Nano with a green dot.

Hardware requirements (Chrome enforced): 22 GB free disk, 16 GB+ RAM recommended, unmetered connection for the initial model download.

Microsoft Edge (Dev / Canary recommended)

Edge uses the same Prompt API standard with Phi-4-mini as the on-device model.

1. Use Edge Dev or Canary with the Prompt API flag enabled:

   edge://flags/#prompt-api-for-gemini-nano
   

   (Edge may label this Prompt API for on-device language model.)

2. Enable the on-device model flag (same as Chrome):

   edge://flags/#optimization-guide-on-device-model
   

   Select Enabled BypassPerfRequirement.

3. Relaunch Edge and verify at edge://on-device-internals.

4. Reload ScrollCollector — header should show Phi-4-mini with a green dot.

Smoke test (service worker console at edge://extensions → ScrollCollector → Service worker):

await LanguageModel.availability({
  expectedInputs: [{ type: 'text', languages: ['en'] }],
  expectedOutputs: [{ type: 'text', languages: ['en'] }]
});
// Expected: "available"

Brave

Brave does not ship Google's Gemini Nano download, so posts System 1 is unsure of fall back to the Rule-based Classifier there (keyword matching, no flags required). System 1 still runs. It does not download a replacement language model.

---

How to Use

1. Browse: Navigate to any enabled site (LinkedIn, YouTube, Twitter/X, Medium, etc.).
2. Curation:
   • Hover/focus on cards in the side panel.
   • Reclassify category tags inline using the Override popup, or use the Left/Right arrow keys.
   • Favorite items with Space or remove individual items with the × button.
3. Ignored Keywords: Add unwanted words in the settings panel to automatically skip them.
4. Export: Open the Export tab and click Generate HTML Diary to save a local styled copy of your feed.
5. Clear: Click Clear stream in the footer to purge all un-favorited items.

---

Licence

ScrollCollector is source-available, not open source. It is free for personal and other non-commercial use: you may use, change and share it under the PolyForm Noncommercial License 1.0.0. Commercial use needs a licence from the owner: hi@fauzul.com.

Copies you obtained under the earlier MIT declaration (up to and including version 0.11.0) keep that licence. The parts bundled with the extension keep their own licences; see THIRD_PARTY_NOTICES.md.

No external network tracking. By default every decision is made on your device; post text leaves it only if you turn on a cloud provider (TypeSafe Jev or OpenAI).

Contributing

Issues and suggestions are welcome. Outside code contributions are not accepted for now (see CONTRIBUTING.md). Report security problems privately (see SECURITY.md).