Free · On-device · Open source

You should never have to pay
to understand something.

Every day you read things you only half-understand: a term, a claim, a paragraph that quietly assumes you know more than you do. The web has one fix for this, and it's a detour: open a tab, paste into a chatbot, lose your place. Or pay $10 a month for a sidebar that does the pasting for you.

JustClarify's on-device engine and Your LLM option are free. Want a cloud model? Bring your own API key (BYOK), or choose the optional hosted engine: 30 asks free, then $3.99 a month. It never switches you there without you choosing it.

Here's the alternative. Highlight a phrase, or speak about anything in the article.

a-long-article-you're-reading.com

Why the team hit pause before the release

Late in the sprint, the engineers made an unusual call. Rather than cram in one more feature, they chose to refactor the legacy module first, arguing that the technical debt had grown untenable, and was quietly slowing every new change to a crawl.

Nobody outside the team could see it, but each shortcut taken months earlier was now taxing everything built on top of it.

Click the highlighted phrase to select it, like you would while reading.

This demo is scripted. Every answer is pre-written and no AI is called. The real extension generates them live, on the engine you picked.

We think the subscription model is backwards. The intelligence already lives on your machine. Chrome now ships an AI model inside the browser itself, and if that isn't enough you almost certainly already pay for ChatGPT, Claude or Gemini. Either way the answer can cost nothing extra, need no account, and never send a word of what you're reading to a server of ours. So that's what we built. Then we open-sourced all of it, so you don't have to take our word for any of this.

Highlight a phrase, or ask about the page out loud. Read the answer where you stand. Keep going.

01Free where it costs us nothing.

Device · Free foreverYour LLM · Based on your subscriptionBYOK · You pay your provider, not usHosted · 30 asks free, then $3.99/mo

Tools like this charge a subscription because every answer costs them server money. Three of the four ways JustClarify can answer cost us nothing: your computer's own model, the chatbot subscription you already pay for, or your own API key going straight to Anthropic, OpenAI or Google. Those are free, forever, with no account and no credit counter. Only the hosted engine runs on our hardware, so only the hosted engine costs money. You pick; nothing switches to a paid one behind your back.

02Private by architecture, up to the line you draw.

On the on-device engine, what you highlight never touches our servers, because there are no servers in the loop. The model runs on your machine and airplane mode works. Choose the hosted engine and your text goes to our server to be answered, because that is the only way that engine can exist. Fact-checks always send the claim and the article's URL, so an article is checked once and the verdict is served to everyone after you. The engine badge on every answer tells you which of those just happened, and the privacy policy spells out the rest.

03Open source, so trust is optional.

The entire extension is public code. Every answer is badged with the exact engine and model that wrote it, and fact-checks only return a verdict with a source you can click. If we ever break these rules, you'll see it in the diff.

Also, without touching anything

Hold `⇧Shift` and say it instead

Highlighting is one way in. The other is speaking: hold ⇧Shift, say what you want, let go. Releasing the key is what ends the turn. No wake word, no open microphone, and nothing is ever captured unless a key is physically held down. What you get back is not a command list. It is an agent that works your page, one step at a time, looking again between every step.

It does the same things your mouse does. Every spoken verb ends in the exact function a click already reaches, so the two ways in can never drift apart.“Explain this”“Fact-check that”“Define quantitative easing”“Translate it”“Read it to me”
And the things your mouse is slower at. Eighty phrases, all of them things you were going to do anyway: reading aloud, scrolling that you can steer mid-flight, tabs, zoom, media, typing into a box and sending it.“Take me to the pricing bit”“Keep scrolling, wait, back”“Next tab”“Click the sign-up button”“Undo that”
A sentence can be more than one action. “Find their refund policy and read me the bit about shipping” is three, and the page after a click is not the page before it. So it takes one step, looks at what that step produced, and only then decides the next. Up to five steps, counted by the browser, not by the model, because something that trusted a model to stop would have no limit at all.
It sees the page as a numbered list, not as a picture. Before each step the extension collects what is actually on screen and actually interactive, the buttons, links and fields, and hands over that list with the ones nearest your cursor first, so “click this one” has something to point at. Naming a numbered thing is one exact step. Describing a button and hunting for it again is two guesses, and the second one is where these systems usually click the wrong thing.
Most of what you say never reaches a model at all. “Scroll down” is settled by a pattern in microseconds, for nothing, with no network round trip. The model is the fallback for the sentences the patterns miss, never the path everything takes, which is why the common commands answer instantly.
A page cannot talk it into anything. The part that decides what to DO never reads the page's prose, only its link labels and headings, so a sentence buried in an article cannot pose as an instruction. What comes back is a verb chosen from a fixed list of forty-five the extension already owns, plus a number. Nothing the model returns is ever run as code, and anything that reads as destructive stops and asks you first.
"This" means what you'd expect it to mean. Say "explain this" with nothing selected and it resolves the most recent thing you could plausibly have meant (your last highlight, the paragraph under your cursor, the thing you just asked about) as a real selection, so the answer arrives exactly where a click would have put it.
Two microphones, because one of them can be switched off. Speech recognition runs in the page when the site allows it. Plenty of sites don't. A page can disable the microphone outright, and http:// pages have none at all. So the extension can hold one grant on its own origin and record there instead, which no website can veto. You grant it once and never see a permission prompt again.
Voice needs the hosted engine. Paid engine Turning a sentence into the right action is a harder job than explaining a paragraph, and it's the one part that doesn't run on Chrome's built-in model yet. Explaining, defining and fact-checking stay free on your own hardware.

For the curious

How a highlight becomes an answer

No magic, and no hidden backend. The one server we do run is named below. Here is the actual pipeline, straight from the source code:

It reads around your selection, not just your selection. When you highlight, the content script walks outward through the surrounding text and captures a semantic window: about two full sentences on each side for Explain, up to six for Expand. That's why answers fit the article you're in, not a dictionary's idea of the phrase.
The prompt is assembled on the page. Your selection, the surrounding passage, the page title, the action you picked, and your density setting are built into a single prompt locally, inside the tab. At this point, nothing has been sent anywhere.
Then you choose who answers it. Four ways to answer, picked in the popup, never switched behind your back:DeviceFreeChrome's built-in Prompt API, Gemini Nano, running on your own hardware. Chrome downloads that model once; JustClarify asks first and shows the progress rather than hanging on a silent spinner. After that, answers stream in with no network round-trip at all.~4GB one-time download~16GB RAM22GB freeWorks offlineYour LLMBased on your subscriptionThe chat subscription you already pay for. JustClarify opens one tab of its own, parked in a collapsed group at the edge of the tab strip, asks ChatGPT, Claude or Gemini there, and reads the answer back. It never touches a tab you opened or a conversation you were having, and it costs nothing beyond the plan you already hold.ChatGPTClaudeGeminiNo extra billBYOKYou pay your provider, not usBring your own API key. Paste a key from Anthropic, OpenAI, Google, Hugging Face or an AI Gateway, and the request goes straight from the extension to that company. The key is stored on your device, no JustClarify server sits in the middle, and we never see it or bill you for anything.AnthropicOpenAIGoogle GeminiHugging FaceAI GatewayHosted30 asks free, then $3.99/moOur model on our server, for machines that can't run a local one and for people who would rather not manage a key at all. This is the only one of the four where your text leaves your machine for us, and the only one that costs money.No setupVoice control included
Definitions don't go to a model at all. Highlight a single word and Define queries a real dictionary (the free, keyless Dictionary API) for the actual entry: part of speech, senses, and the dictionary's own example sentence. A definition is a lookup, not a generation. Only words with no entry (jargon, proper nouns, coinages) fall back to a contextual explanation.
BYOK: your key, your models, your bill. Flip on "use your own AI key" in settings, paste a key from Anthropic, OpenAI, Google, Hugging Face or an AI Gateway, and the same prompt goes straight from the extension to that company. JustClarify reads the prefix to know whose key it is and picks a sensible model unless you name one. The key is stored on your device and sent only to the company it belongs to. No JustClarify server sits in between, we never see the key, and we never charge you a cent for using it.
Without a key, the hosted model. It is a real server. 30 asks free, then $3.99/mo Machines that can't run Gemini Nano used to install this and get nothing. Now the prompt goes to justclarify.xyz, which holds our API key server-side (a key shipped inside an extension is a zip file anyone can unpack) and streams the answer back. It is last in the chain because it is the only tier that costs us money and the only one where your text leaves your machine for us.
Fact-checks are shared, on purpose. A claim and the page's URL go to our cache, which checks that article against published rulings once and serves the result to everyone who reads it after you. It's the one part of the product that gets better because other people used it, and the one place a URL you visited is stored. Verdicts come back with a source you can click, or they don't come back at all.
The badge never lies. Every answer is stamped with the engine and the model that produced it, so a reply from your own hardware, from your chat subscription and from our server can never be mistaken for each other. If you want to know where a sentence came from, it's written on the sentence.

The web ships one version of every page to everyone. JustClarify recompiles it for what you already know.

Add it to Chrome · Free Read the source

You should never have to pay to understand something.