RSSgate

Most feeds hand you a headline and two sentences, then send you to a page full of ads to read the rest. RSSgate goes and gets the rest. For every new post it fetches the article, asks a language model to write a short, faithful digest of it, and puts the result in one quiet stream you can read top to bottom.

An example card, made up for this page. Real cards look like this: the feed, the time, its categories, a digest written from the full article, and a link back to the source.

How it works

You add feeds in the admin panel. RSS and Atom work as you'd expect, and so do ordinary web pages with no feed at all: paste a homepage and RSSgate looks for a real feed first and checks that it actually parses. If there isn't one, it offers to follow the page itself, picking the article links out of it.

When a new post arrives, RSSgate downloads the linked page, pulls out the article text, and sends that to the model with a prompt asking for the important facts, names, numbers and quotes, and nothing else. Each article is digested once. If the same text shows up again, from a reposted item or a syndicated copy, the earlier digest is reused at no cost.

The reader is a single stream, newest first. A post counts as read once it has actually been on your screen for a few seconds, or once you scroll past the whole thing, or when you tap it. Sidebar counts tell you what's left in each feed, and categories from the feeds themselves can be combined into filters.

Your model, your machine

It talks to anything that speaks the OpenAI chat API: vLLM, llama.cpp, Ollama, LM Studio and the like on your own hardware, or OpenAI, OpenRouter and Anthropic if you'd rather pay per token. Digesting is plain extraction work, so on local reasoning models RSSgate turns thinking off by default. On one home server that took a typical digest from 51.8 seconds and 3,127 output tokens to 7.4 seconds and 405, with the same result.

Images are downloaded once and served by your own instance, so reading a post never pings an ad network. The admin panel keeps a running count of tokens spent today, this month and in total.

If the model's server goes down for maintenance, posts aren't marked as failed. They wait in the queue, RSSgate checks back on its own, and digesting picks up where it stopped.

On your phone

RSSgate is built to live on a phone's home screen. In Safari, open your instance and choose Share → Add to Home Screen: you get its own icon, pull-to-refresh, and article links that open in Safari while the reader keeps your place. Over HTTPS it also reads offline: the stream and the photos you've already seen stay available with no signal, and anything you read meanwhile syncs when the server answers again.

Reach it over a VPN rather than opening a port. With Tailscale, one command on the machine running RSSgate gives you an HTTPS address on your tailnet, and HTTPS is what lets iOS offer the share sheet and clipboard:

tailscale serve --bg 8088     # then open https://<machine>.<tailnet>.ts.net

Any post can be shared as an image, either a tall card sized for phones or a wider one, with a QR code that leads back to the original article.

On your desktop

The same trick works on a computer, and it is worth doing: the reader opens in its own window with its own icon and no address bar, while article links still open in your normal browser. In Chrome, Edge or Brave use the Install app icon in the address bar (Edge also has Apps → Install this site as an app); in Safari on macOS, File → Add to Dock. Each app window keeps its own scroll position, and pull-to-refresh works there too.

One quirk: an installed window sometimes reports its own dark-mode state instead of your system's, so "follow the browser" can disagree with the rest of your desktop. Pin it from the admin panel under Display & sharing → Colour scheme: auto, always light, or always dark, for every page and device.

Being a good guest

A reader that fetches every article is a heavier visitor than one that only reads feeds, so RSSgate spaces out requests to each site, and backs off for an hour (doubling on repeat) when a site answers too many requests or forbidden. Posts wait rather than fail.

Some sites put a bot check in front of their articles. RSSgate is upfront about how it handles that: it presents itself as a browser. It sends a browser User-Agent and keeps each site's cookies. If a check still appears, it retries once with full browser headers, and then opens the page in headless Chromium with a persistent profile and the automation markers hidden. Cookie walls are dismissed by simple rules, and if the page is still stuck the model may click a button that is actually on the page, wait, or give up. It can never type or submit anything. The pacing rules still apply, and the browser opens at most once per site every ten minutes.

When none of that gets through, RSSgate shows the post from the text the feed itself carried, marked as an excerpt, and tries the page again later. If a site keeps checking, it stops asking for pages and probes once a day, and the admin panel tells you which feeds are affected.

Install

git clone https://github.com/evulnet/RSSgate && cd RSSgate
python -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/python -m rssgate    # http://127.0.0.1:8088

# optional: the headless-browser step for bot-checked sites
.venv/bin/pip install playwright && .venv/bin/python -m playwright install chromium

Everything is configured from one YAML file, and almost all of it can be changed from the admin panel while it runs. Settings save as you change them.

RSSgate has no login. It listens on 127.0.0.1 by default and it should stay that way: the VPN route above reaches it without exposing the admin panel, or the model you're paying for, to the rest of your network.

Runs asone Python process, Flask and SQLite
Configured byone YAML file and the admin panel
ModelsOpenAI-compatible servers, OpenAI, OpenRouter, Anthropic
Tested byan offline suite and a real-browser suite