Either it quotes the page
A supported claim carries the sentence it came from, word for word, and clicking it scrolls the real page to that sentence and paints a highlight over it. You are checking the source, not a second summary of it.
Chrome extension · Free on your own key
Briefly summarizes the article you’re reading into short claims. Click any claim to see the exact sentence it came from. Anything it can’t trace back is marked Unsupported — not hidden, not dropped.
Briefly
journal.example.com/sleep-reaction-time
What Briefly says
What the page says
Participants who slept fewer than six hours showed a 14% slower mean reaction time the following morning.
The effect held across all four cohorts, including the group that reported feeling well rested.
Caffeine intake was not measured, which the authors list as the study’s main limitation.
Nothing in this page says anything about long-term decline. That is why the third claim carries a mark instead of a quote.
The one rule
Every summariser writes confident sentences. The one you cannot catch is the confident sentence the article never supported — catching it would mean reading the article, which is the thing you were trying to avoid. So Briefly makes a narrower promise, and one you can check in ten seconds.
A supported claim carries the sentence it came from, word for word, and clicking it scrolls the real page to that sentence and paints a highlight over it. You are checking the source, not a second summary of it.
If the sentence the model cited is not in the page, the claim stays on screen carrying the word Unsupported, an icon, a solid edge and a tinted ground — four ways to read the same fact, so removing any three still leaves it legible.
No “partially supported”, no paraphrase that counts as a quote, no quietly dropping the ones that failed so the list looks clean. A summary with the failures removed is a summary that lies by omission.
Briefly compares each cited sentence against the text your own browser extracted — the model is never asked whether it was right. So a claim produced on a cheap open-weights model carries exactly the same proof as one produced on the most expensive model in the list. A weaker model gives you fewer supported claims. It cannot give you one that looks supported and is not.
Worth being exact about: Briefly verifies that the quoted sentence is genuinely in the page. It does not verify that the page is telling the truth. If an article is wrong, a faithful summary of it is wrong too — what Briefly removes is the layer of error the summariser adds on top.
One click, four steps
Briefly has no content script and no background activity. Until you click the icon it is not running on the page you are reading — there is nothing to switch off, because nothing was on.
That click is what grants access to the tab, for that tab, at that moment. Chrome
calls this activeTab, and it is why Briefly asks for
no site access when you install it.
The page text is extracted locally. If there is no readable article, or the page is short enough to simply read, Briefly says so and stops there — a correct outcome that costs nothing.
The article text goes to the provider you chose, on your key, and comes back as a list of claims each carrying the sentence it was drawn from. This is the only network request Briefly makes.
Each cited sentence is matched against the text your browser extracted. Found is Supported. Not found is Unsupported, on screen, with the reason. Then you read three claims instead of three thousand words.
Bring your own key
Most summarisers are free because you are paying in some other currency. Briefly is free because the inference bill is yours — which also means your page text goes from your browser straight to the provider you picked, under your own agreement with them. Briefly’s server never sees it.
The one most people already have a key for.
Free tierGPT models, billed to your own OpenAI account.
Claude models, on your Anthropic console key.
One key, nearly every model on this page. Several cost nothing.
Free tierThe fastest option here, with a generous free rate limit.
Free tierVery cheap per summary, billed to your own balance.
European hosting, with a free experiment tier.
Free tierGrok models on an xAI console key.
Open-weights models, very fast, daily free allowance.
Free tierA broad open-weights catalogue on one key.
Hosted open models with routed endpoints.
The long tail of open-weights models, cheaply.
Open models with an hourly and daily free allowance.
Free tierThe model field takes anything you type and sends it as written, so a model released after this build still works.
Setup, once
Pick a provider, open their keys page, copy the key, paste it into Briefly’s Settings and tick the box. The setup guide walks through it provider by provider, including which ones cost nothing and what every error message means.
What leaves the browser
Privacy claims are worth exactly as much as the reader’s ability to test them. Each of these is visible from Chrome’s own dialogs or its Network panel, with no need to take anybody’s word for it.
Briefly declares no host permissions, so Chrome shows no “read your data on all websites” warning when you add it. Access to a provider is requested one at a time, in the click where you configure it, and you can withdraw it from Chrome’s own extension settings. Check it in the install dialog.
There is no content script and no background polling. Open the Network panel, browse for an hour, and Briefly appears in it zero times until the moment you use it. Check it in the Network panel.
On the free tier there is exactly one recipient and it is the provider you chose. Your key is never sent to Briefly, and Briefly’s server is never in the path of your page text. Check it in the Network panel.
Before the first summary, Briefly tells you who is about to receive the text — by name, the provider you actually configured, not a generic “third-party AI service”. Consent to send your reading to an unnamed party is not consent. Check it on first use.
By decision, not by omission
Each of these would make some screen look better and some number go up. They are listed here so that if one of them ever appears, you know it was a change of position rather than a feature.
Dropping the failures would make every summary look clean. It would also make the mark meaningless, because you could no longer tell a careful summary from a filtered one.
When an article is too long to send whole, Briefly says which share of it the summary covers. Presenting the first third as the whole would be a false claim about the document, made by us rather than by the model.
No content script, no pre-fetching the page you are “probably about to” summarise. Anything that reads ahead to feel fast is reading pages you never asked about.
There is no event pipeline, no session recording and no list of what you have read. It costs us the product analytics everybody else has, and that is the trade.
On Pro, the request that carries page text carries no account identifier, and the one that identifies your licence carries no page text. They are separate requests on purpose, so neither can be joined to the other.
Rewrite a claim in the editor and its source quote is removed automatically. A real sentence from the article underneath text that no longer says what the article said is the most convincing way to be wrong.
What the money buys
Every feature is in the free version, with no limits — the anchoring, the unsupported marking, translation, the editor, unlimited saved summaries and export. None of it sits behind the price. The single difference is whose API key runs the model.
The complete product, running on a key you fetch once. Nothing is held back and nothing expires.
$0on your own API key
Everything in Free — which is everything — with the setup step removed. Pick whichever way of paying suits you.
Straight answers
It means the sentence the claim is built on is genuinely in the page, word for word. Briefly finds that sentence in the text your browser extracted, and clicking the claim scrolls to it.
It does not mean the claim is true in the world. Briefly checks the summary against the article; it does not check the article against reality.
Because running the model costs real money, and that cost has to be paid by somebody. Briefly asks you to pay it directly to a provider of your choosing, rather than paying it with your reading history.
It has a second effect that is arguably the bigger one: with your own key, the page text never touches Briefly’s server at all. If you would rather not deal with keys, that is exactly what Pro is for.
Groq is the usual answer — free, fast, and a generous rate limit. OpenRouter is the other good one, because a single key reaches nearly every model in the list and several are free.
Google’s free tier also works, with one condition worth knowing about first: see choosing a provider in the setup guide.
On a free tier, nothing. On a paid one it is a fraction of a cent for a typical article on the cheap fast models Briefly recommends — the exact figure is your provider’s price list, and Briefly shows each model’s input and output price beside its name when you pick one.
Cheap and fast is genuinely the right choice here. The checking is done by your own machine, not by the model, so paying for a larger model buys better writing rather than better verification.
On the free tier, no — there is one request and it goes to the provider you configured. On Pro, the request carrying your page text carries no identifier of you, and the request that identifies your licence carries no page text.
Either way there is no analytics, no browsing history and no list of what you have summarised anywhere but your own device.
Briefly says there is no readable article, and sends nothing. Same for a page short enough that reading it is quicker than summarising it — it tells you the word count and stops.
Chrome’s built-in PDF viewer is not an HTML document, so it comes back as no readable article too.
Yes. The model field is free text and sends whatever you type, so a model released after this version of Briefly still works. The list is there to help you find one, not to decide which may run.
One requirement: the model has to support structured JSON output. Briefly asks for a JSON object of claims, and a model that answers in prose gives nothing usable. That is why the built-in list only contains models that promise it.
Settings has a control that removes the stored key and withdraws the host permission in the same click. Chrome’s own extension settings will also revoke site access at any time, without uninstalling anything.
Your saved summaries are not touched by either, and stay on your device until you clear them.
Feedback
A provider you would like added, a summary that did not read the way you expected, or something that could simply work more neatly — send it over. It goes straight to the person building this.