Headline GatorSettings

News alerts from the major newsrooms, in one place

What this is

News apps send breaking-news alerts to your phone. Each one is an editor deciding that something matters enough to interrupt you. Headline Gator collects those alerts from 26 news apps and puts them in one list, newest first, so you can see what every newsroom thought was urgent without installing 26 apps or letting any of them buzz your phone.

It is a reading list, not a newspaper. Nothing here is written, edited or ranked by us: the wording is the publisher’s own alert text, and the order is simply the order the alerts arrived.

How it works

A dedicated Android phone sits on a shelf with the news apps installed and their alerts switched on. It does nothing else. When an app posts a notification, an automation tool on the phone forwards the publisher, the alert text and the timestamp to this site, which stores one row per alert.

The awkward part is the link. A push notification contains no web address — that was the first thing we checked, by dumping real notifications. The destination lives inside Android’s tap-handling data, which apps are not required to expose, so the alert arrives as a headline with nowhere to go. The site then has to work out which article it announced:

About 70% of settled alerts end up with a direct link. The rest keep their Search link, which is a fair description of what the site knows.

The model that reads the rewrites

The matching question is small and strange: given one alert and eight candidate articles, which one — if any — does this alert announce? It needs judgement, because “markets tumble as the Fed holds” and “Fed leaves rates unchanged, stocks slide” are the same story in different words. It does not need eloquence, and it is asked thousands of times a month.

So the site uses Jev, a small model built for exactly this shape of problem, through OpenRouter’s decisions interface. Rather than writing an answer in prose that then has to be parsed, it returns a typed choice — one of the numbered articles, or none — along with a calibrated confidence: a number that means what it says, so 0.9 really is more reliable than 0.7. The site ignores anything under 0.6 and leaves the alert unlinked instead.

That last part matters more than the hit rate. A wrong link is worse than no link, because it quietly sends you to the wrong article, and the most tempting mistake here is a sibling — the follow-up or the explainer about the same event rather than the piece the alert announced. Before adopting it we ran it against alerts whose correct articles we already knew, and separately against the same alerts with the right answer removed, to see whether it would invent a match rather than decline. It answered “none” when it should, and picked no wrong article at any confidence.

In practice it answers in well under a second for a fraction of a cent, and it is responsible for 30% of every link on this site — the alerts that plain text matching could never have caught. Its accepted picks average 0.85 confidence.

How to use it

What about privacy

There are no accounts, no passwords and no email addresses. Your edition and source choices live in cookies on your device; a profile is a list of source names and nothing else. Clicks on stories are counted in aggregate so we can see which sources earn their place, without IP addresses or anything that identifies a visitor.

Why is a story missing or wrong

The usual reasons: the publisher never sent that alert to Android (alerts differ by platform and by account); the alert arrived but its article isn’t in any public feed yet, so it is still hunting for a link; or the phone itself stopped forwarding. That last one is visible — if the phone goes quiet, a notice appears at the top of the list rather than leaving you to wonder why the news stopped.

Send feedback · Back to preferences