I Built an AI-Powered RSS Aggregator Without Writing a Line of Code
Remember Google Reader? When it shut down in 2013, I lost my favorite way to keep up with technology news. I tried other RSS readers over the years, but they all had the same problem: by the end of the day, I had hundreds of unread articles and no good way to find the ones that actually mattered to me.
So I built News Diet, an AI-powered RSS aggregator that uses local machine learning to score and filter articles based on my interests. The interesting part is that I did not write a single line of code myself. Everything was “vibe coded” using AI assistants. This project shows how vibe coding can help you build useful personal tools quickly when production-grade quality is not needed.
What News Diet Does
News Diet fetches articles from your RSS feeds and analyzes them using a local language model running through Ollama. Each article gets a relevance score from zero to ten based on topics you care about. The system extracts topic tags that match your configured interests, assesses article quality, and calculates a score based on how many of your topics appear in the content. Articles with zero matching topics score low, while articles covering multiple interests score high. You can set a minimum relevance threshold, and anything below it is hidden automatically. The AI also generates a short summary for each article so you can decide what to read without opening multiple tabs.
How It Works
The project runs with Docker Compose and includes a FastAPI backend, MongoDB for storage, and Ollama for the local language model. I chose the qwen2.5:3b model because it runs on a CPU without needing a GPU, making it perfect for a home server or mini PC. The application fetches new articles every hour, processes them through the AI model, and stores everything locally. Your data never leaves your machine.
The interface uses HTMX for dynamic updates and the Catppuccin color scheme with light and dark modes. You can manage feeds through a web interface, add topics of interest or exclusions, star articles, and mark things as read. Everything is designed to be simple and self-contained.
Why Vibe Coding Works for Personal Projects
When you build something for yourself that runs in your own network, the usual software development rules do not really apply. You do not need perfect test coverage, security audits, or long-term maintenance plans if you are the only user. This is where vibe coding shines. I described what I wanted, the AI generated the code, and I kept iterating until it worked. The entire project came together in a few sessions without me writing any code manually.
For personal tools like this, the trade-offs of vibe coding matter less. The code might not follow every best practice, and there might be unhandled edge cases. But if the tool does what you need and you can fix issues with more AI assistance, that is good enough. The alternative is spending weeks learning a new framework and writing everything by hand for a project that only you will use.
Try It Yourself
The project is open source and available at github.com/miguelangelnieto/news-diet. If you want your own AI-powered RSS reader, you can clone the repository and run it with Docker Compose. The setup is straightforward, and the README includes detailed instructions for different configurations. You will need at least eight gigabytes of RAM and a decent CPU, but no GPU is required. The model downloads automatically on first run, and then everything works offline.
I am hosting my instance privately because this is a personal tool, but there is no reason you cannot run your own. If you have been missing Google Reader or want a better way to filter your information diet, give it a try. The project is released under the MIT license, so you can modify it however you want. And if vibe coding can build something like this for me, imagine what it can do for your own personal projects.