Getting Briefly working on the free tier

1. Install Briefly

Briefly is not on the Chrome Web Store yet. When the listing goes live, installing it becomes one click from the home page and this step becomes a sentence long. Until then there is nothing to install, and the rest of this guide is what to do the day there is.

The one thing worth noticing at install time is what Chrome does not say. Briefly declares no site access, so there is no “Read and change all your data on the websites you visit” line in the dialog. It cannot reach any page, or any network, until you give it both — one page at a time, one provider at a time.

2. Why Briefly asks for a key at all

Running a language model over an article costs money every time. Somebody pays it. Briefly’s free tier asks you to pay it directly to a provider you choose, which is why there is no advertising, no tracking and no data sale funding it instead.

There are two consequences, and they are both good ones:

• Your page text never reaches Briefly. The request goes from your browser straight to the provider, under the agreement you have with them. There is no server of ours in the path to trust.
• You choose the model. 484 models across 13 providers. If you want the cheapest thing that works, or the fastest, or a specific open-weights model, that is your decision rather than ours.

The cost is this page: a couple of minutes of setup before Briefly does anything. If you would rather skip it entirely, that is exactly what Briefly Pro sells — we run the model, you never see a key. Every feature is identical either way.

3. Choose a provider

A provider is a company that will run a model for you and bill you for it. You need an account with exactly one of them.

All thirteen, with what matters about each

The one warning on this page

This is also the general shape of the question for every provider: what happens to the text after it arrives is decided by their terms, not ours, because it is your account and your agreement with them. Briefly’s job is to name the recipient clearly and send it nowhere else. Reading a provider’s data-use page once, before you paste the key, is five minutes well spent.

Choosing a model, once you have a provider

Pick something cheap and fast. This is counter-intuitive and it is genuinely correct: the verification is done by your own browser, comparing quoted sentences against the article text, and the model has no part in it. A larger model writes better claims. It cannot produce a claim that looks verified and is not, and a smaller one cannot either.

The model must support structured JSON output — Briefly asks for a JSON object of claims. Every model in Briefly’s built-in list already promises this, which is why the list is shorter than the provider’s full catalogue. The field also accepts anything you type, so a model released after your version of Briefly works fine; it just has to meet that one requirement.

4. Get your key

Every provider follows the same pattern: make an account, open the API keys page, create a key, copy it. Three of them in full, then the general version.

Groq — free, and the fastest route through this page

1. Open console.groq.com/keys and sign in. A Google or GitHub account works; there is no card to enter.
2. Select Create API Key and give it a name such as "Briefly".
3. Copy the key. It begins gsk_. Groq shows it once and will not show it again — if you lose it, delete that key and make another.
4. Suggested model: openai/gpt-oss-20b, which is what Briefly fills in for you.

OpenRouter — one key, 286 models

1. Open openrouter.ai/keys and sign in.
2. Select Create Key, name it, and copy it. It begins sk-or-.
3. Optionally add a few dollars of credit. You do not have to: models whose name ends in :free cost nothing, and Briefly’s default for OpenRouter is openrouter/free.
4. Later, if you want a different model, search the picker in Briefly’s Settings rather than going back to OpenRouter — the same key already works for all of them.

Google Gemini — if you already have an account

1. Open aistudio.google.com/apikey and sign in.
2. Select Create API key, and pick or create a Google Cloud project for it.
3. Enable billing on that project if you would rather your reading was not used as training data. This is the warning above, and this is the step that answers it.
4. Copy the key. It begins AIza. Suggested model: gemini-3.1-flash-lite.

Any of the other ten

1. Pick the provider in Briefly’s Settings first.
2. Use the Get an API key from … link Briefly shows directly under the field. It goes to that provider’s own keys page, so there is no guessing which of several dashboards is the right one.
3. Create a key and copy it.
4. Check the shape against the hint beside the link — Briefly tells you what that provider’s keys start with, and checks it before anything is sent.

5. Put the key into Briefly

Everything typed in Briefly lives in the full tab rather than the popup, and this is deliberate: Chrome destroys a popup the moment it loses focus, so a half-pasted key would vanish without a word.

1. Click the Briefly icon, then Settings. Briefly opens its own tab.
2. Find the section called Your model. If it is not there, this install is on Pro — Briefly runs the model for you and there is nothing to configure.
3. Pick your Provider. The list is grouped into Model makers — companies that train their own — and Many models, one key, which run everybody else’s. Each shows how many models it carries, and a note appears underneath if it has a free tier.
4. Choose a Model. Type to search the catalogue by name or id; each result shows its context window and its price. You can also type a model id that is not in the list and Briefly will send it as written.
5. Paste your key into API key. The field is masked, so it is not readable over a shoulder or in a screen share.
6. Tick the consent box. It says that page text will go to this provider under your own API agreement, and that they — not Briefly — decide what happens to it. The button will not submit without it.
7. Select Save and use this model.

6. The prompt Chrome shows you

Saving triggers a Chrome permission dialog naming one host — the provider you just picked, and only that one. Accepting it is what lets Briefly reach them at all.

This is the permission Briefly deliberately did not ask for at install. It is requested here because here is the first moment it is actually needed, and it is requested one provider at a time: choosing Google never grants the other twelve.

If you decline, Briefly says “Briefly cannot reach <provider> without that permission. Nothing was saved.” and your key is not stored. Nothing is half-configured; run through step 5 again when you are ready.

7. Your first summary

Open an article — a real one, a few hundred words or more — and click the Briefly icon.

The first time, a consent screen appears naming the provider that is about to receive the text. It names the one you configured, by name. Accept it once and it does not come back.

Then one of these happens, and all of them are normal:

• A list of claims. The ordinary case. See the next section.
• “There is no readable article here. Nothing was sent.” The page is a feed, an application, a search result or a PDF viewer rather than an article. No request was made and nothing was spent.
• “This page is N words — short enough to read. Nothing was sent.” Summarising it would take longer than reading it.
• “This covers the first N% of the document.” The article was too long to send whole. You get claims for the part that was sent, and the share is stated rather than hidden.
• “The model is unreachable, so these are sentences taken straight from the page.” The provider could not be reached, so rather than nothing, Briefly falls back to sentences lifted verbatim out of the article. No model wrote them, which is why they are exactly the page’s own words.
• An error naming your provider. Go to troubleshooting; each message says which side the problem is on.

8. Reading the result

Every claim carries one of exactly two labels. There is no third state and no in-between.

• Supported — the sentence this claim was drawn from is genuinely in the page, word for word. Click the claim and the page scrolls to that sentence with a highlight over it. That is the check; do it on anything that matters.
• Unsupported — the sentence the model cited is not in the page, or it gave no sentence at all. The claim stays on screen, marked with the word, an icon, a solid edge and a tinted background, and a line underneath saying which of the two happened.

An unsupported claim is not a crash and not a bug report. It is the product doing the one thing it exists to do: the model produced something the article does not say, and you are being told so instead of being handed it as fact.

A run with several unsupported claims usually means the model is too small for the article, or the page is a mixture of article and interface. Trying a slightly stronger model is the useful response.

9. Translate, edit, save and export

All of this works the same on the free tier as on Pro, with no limits. On free it spends your own key, exactly like a summary does.

Translate

Translates the claims while each one keeps its link to the original source sentence, so a translated claim still jumps to the English — or whatever the page is written in — sentence it came from. The proof is not translated, because a translated quotation is no longer a quotation.

The editor

The full tab has an editor for turning claims into something you would actually post. One rule is enforced automatically: editing a claim removes its source quote, and the claim is labelled as edited wherever it appears afterwards. A real sentence from the article sitting underneath words that no longer say what the article said is the most convincing possible way to be wrong.

Saving and export

Saved summaries are unlimited on both tiers and stored on your device. Markdown export is likewise on both. Re-saving a page you already saved moves it to the top rather than storing it twice, and nothing is dropped to make room.

10. What it costs

On a free tier: nothing, until you hit that provider’s rate limit, at which point it starts working again on its own schedule.

On a paid tier: a typical article is a few thousand tokens in and a few hundred out. On the cheap fast models Briefly suggests, that is a fraction of a cent per summary. The exact number is your provider’s price list, and Briefly shows each model’s input and output price in the picker so you can compare before choosing.

Three things keep it near zero:

• Briefly only runs when you click it. There is no background summarising, so you are never billed for a page you did not ask about.
• Pages that are not articles cost nothing. Extraction happens locally and short or unreadable pages stop there, before any request.
• A cheap model is the right model. The verification is local, so the expensive model buys you better prose and nothing else.

Set a spending cap in your provider’s dashboard if they offer one. It is good practice for any API key, and nothing in Briefly can exceed it.

11. Troubleshooting

Briefly names the provider in every failure, and says explicitly when the problem is your account rather than ours.

12. Changing or removing a key

To change provider or model: go back to Settings → Your model, pick the new one and save. Changing provider resets the model field to that provider’s suggested one, because carrying an OpenAI model id across to Anthropic produces an error in a field you never touched.

To remove the key entirely: use the remove control in the same section. It deletes the stored key and withdraws the host permission in one action, so Briefly can no longer reach that provider at all.

From outside Briefly: Chrome’s extension settings will revoke site access at any time without uninstalling anything, and deleting the key in your provider’s dashboard makes the copy Briefly holds useless immediately. Either is a complete answer.

Your saved summaries are not affected by any of this. They live on your device and stay until you clear them yourself.