How it actually works.

What you do

You paste a URL into the box. We open the box. The box accepts anything that has text on it, including articles, blog posts, research papers, long threads, newsletter archives, podcast transcripts, documentation, and the occasional PDF you found at the bottom of a government website.

We do not accept videos. We do not accept images of text. We do not accept links that require a password, though we will handle paywalls that hand out a preview — we read what the paywall hands us and tell you honestly that the rest was not available.

What we do

We fetch the page with a respectful user agent. We follow redirects. We wait a reasonable number of seconds for the page to finish loading. If the page is behind a cookie wall, a GDPR banner, or an email capture overlay, we read through it the same way a human reader would — by ignoring the overlay and looking at the text underneath.

We strip away the navigation. We strip away the footer. We strip away the fourteen unrelated newsletters the publisher tried to sell you. What remains is the text the author actually wrote, which is, in almost every case, roughly a tenth of the file you loaded.

We feed that cleaned text to a language model we tune for this job and nothing else. The model does three passes. The first pass writes a hundred and fifty words that cover the piece as a whole. The second pass pulls five specific claims that carry the argument. The third pass finds one sentence worth remembering — usually not the one the author thought was the best, but the one that survives without the rest of the piece around it.

What comes back

A single structured object. Title. Author if we can find one. Date if we can find one. A paragraph. Five bullets. One quote. A source link back to the original, which we never hide and never rewrite. That is it.

You can read it on the web. You can copy it as Markdown. You can export it as JSON. Pro users get an endpoint they can call from a shortcut, a CLI, a Raycast extension, or whatever reading tool they happen to use. The API returns the same object as the web, so you can switch between the two without rewriting anything.

What we do not do

We do not write opinions. The summaries are descriptive. If the author argues that the moon is made of cheese, we write that the author argues that the moon is made of cheese. We do not fact-check. We do not contradict. We do not add our own views.

We do not pretend to have read what we could not access. If the page required a login we could not get through, we say so. If the page was a video, we say so. If the page was empty, we say so. A summary of nothing is not a summary worth sending.

We do not train models on what you save. We do not sell what you save. We do not build a graph of what you read. The library is yours.

What to expect on accuracy

The summaries are not perfect. They are closer to what a careful friend would tell you after reading the piece than to a court transcript. For casual reading, they are reliable. For citation, they are not a substitute for the original. We make that distinction clear inside the app too, next to the quote, every time.

If we get something wrong, there is a small button that sends the bad summary to us. We read all of them. If it is a pattern, we retune. If it is a one-off, we flag the URL and move on. Either way we tell you what happened.

What it costs to run

Every summary has a real cost — roughly half a cent in compute. The free tier exists because most people read ten things a day and never come back, and that is fine with us. The Pro tier covers the users who read a hundred things a day. That is the whole economic model. It is not a lot more interesting than that.

Questions we get

Does it work on paywalled articles?

It works on the parts that the paywall lets us see. If a publisher gives you the first three paragraphs for free, we summarize those three paragraphs and label the result as partial. We do not try to bypass paywalls. That is the publisher's content, not ours.

Does it work in languages other than English?

Yes. The summary comes back in the language of the source by default. You can also ask for the summary in your own language, which we do on the fly. The structure and the accuracy are the same.

Can I use it inside my team?

The Pro plan is per seat. If you want something bigger or custom, send us a note through the contact form. We are three people, so we will answer quickly, but we will not sell you an enterprise contract. There is no enterprise tier. We are honest about that.

Will you ever add comments, likes, sharing?

No. aisummary is a reading tool. It is not a social product. We do not plan to make it one.