Rebuilding a Simpler, Local-First Video Aggregator

2 min read #Web dev #PWA #IndexedDB #Cloudflare

I rebuilt my YouTube content aggregator and filter (video.vamonossoftware.com ) as a streamlined, local-first Progressive Web App (PWA). By stripping out over-engineered features like local LLM tagging and shifting to a client-side IndexedDB model, the new version establishes a fast, responsive, and maintainable foundation.

The Architecture

The application operates entirely client-side, eliminating remote database dependencies and hosting costs.

graph TD A[Remote Source URL / GitHub] -->|Import & Sync| B[Sources Config] C[YouTube API Key] -.->|Direct API Query| D[YouTube Feed Data] E[Cloudflare Worker RSS Proxy] -.->|Fallback Fetch| D B --> D D --> F[Filter Rules: Include / Exclude] F --> G[Topics Feed: Motorsport, Tennis, etc.] G --> H[(Browser IndexedDB)]

1. Local-First Storage

All feed data, subscription lists, topics, and filtering rules are stored directly in the browser using IndexedDB. This keeps interactions instantaneous and respects user privacy.

2. Hybrid Feed Fetching

To bypass CORS restrictions and API limits, the aggregator uses a two-tiered fetching strategy:

  • API Key Mode: If you provide a YouTube API key in the settings, the app queries YouTube’s data API directly from the browser.
  • RSS Fallback: Without an API key, the app requests YouTube channel RSS feeds through a lightweight proxy deployed as a Cloudflare Worker.

3. Shareable Sources & Topic Filters

Subscriptions are structured under Sources. In addition to local configuration, sources can be imported and synced from remote URLs (e.g., a raw JSON file hosted in a GitHub repository). This allows users to curate and share subscription feeds.

Customizable Include/Exclude rules then process these feeds, sorting videos chronologically under user-defined Topics (such as Motorsport or Tennis).

Lessons in Simplicity

The first iteration of this aggregator was overbuilt. It attempted to auto-tag videos using a local LLM, introducing unnecessary latency, configuration friction, and complexity.

By replacing the LLM with straightforward string-matching filters (Include/Exclude rules), the app became faster and easier to maintain. This simpler foundation is more reliable and makes it easier to identify which features are actually worth adding next.

The AI-Assisted Build Stack

Developing this iteration was an exercise in high-velocity prototyping using modern AI tooling:

  • Claude Design & Claude Code: Handled UI/UX generation, code refactoring, and test suite execution.
  • OpenSpec: Streamlined structural specification.
  • Cloudflare Pages: Auto-deploys the build on merge to main, keeping the deployment loop completely hands-off.

The result is a clean, dependency-light codebase that delivers a predictable, algorithm-free YouTube feed.

Still very much in the alpha phase where I’m trying to figure out how it should work for the best user experience. Use at your own risk…

If you enjoyed this article, feel free to support my work on Ko-fi!

Support on Ko-fi

Related Writing