Press ⌘G (Ctrl+G on Windows). Say what you want.
Pharos is the light that guides you around the web. Ask for the refund rule, the cancel button, the page you need or the tab you lost, in your own words, and it shows you where.
Four hunts, one key. Here is what they replace.
The old way, and the new
⌘F only finds the words the page happens to use. So you scroll, you click through menus and tabs, you copy and paste. Six everyday errands, the old way and with Pharos.
The page says “Delay Repay”. So you read all of the small print.
Lit on the page in about half a second.
The link says “End your contract”, in grey, under everything else.
Ringed on the page. Enter clicks it for you.
Thirty-eight favicons, and you click them one by one.
Every open tab answers. The right one opens.
The receipt is right there. You type it in again anyway.
Copy once, press one key. Seven fields, and the date too.
sk_live_51Hq…
Sent · 09:41
Pasted, sent, and in someone else’s logs before you notice.
Held before it lands. Checked on your machine, nothing sent.
Second post down: the score you were saving for tonight.
Folded by a rule you wrote in plain English.
One key for all of it. Press ⌘G (Ctrl+G on Windows) on any page and say what you want.
Now one of them, start to finish, on a docs page you have never seen.
Find
Type what you mean, the way you would ask a colleague. Pharos ranks every passage, button and field on the page by how likely it is to be what you meant, lights the best one and brings it into view. Enter takes you there.
That was Find. It is how people meet Pharos.
Beyond Find
Once Pharos can read a page, it can do more than point. The same key fills in what you already have, holds back the paste you would regret and quiets the feeds you scroll.
First, everything you type twice.
Fill
A receipt, a booking, an address, a CV, a row of a spreadsheet. Copy it, open the form and press ⌘G. Fill works out which piece belongs in which field, shows where each value came from and how sure it is, and fills the ones you approve. Here, a hotel receipt becomes an expense claim.
Fill reads what you copy. Guard watches what you paste.
Guard
A key from any of 100+ services, a seed phrase, a card number, a one-time code. Before it lands in a chat box or a form, a small chip asks whether you meant to and names what it saw. The checks run in your browser, on the open-source gitleaks and Nosey Parker rules and Guard's own: local by default, with an optional cloud check you switch on. Try it: copy a sample below, then paste it into the message box.
And Rules watches what your feed puts in front of you.
Rules
Write a rule the way you would say it: no spoilers for tonight's match, no LinkedIn cringe. Posts that fall under it fold to one line you can reopen, and the rest of your feed stays as it was. X and LinkedIn as you know them; the posts and the people are made up.
One key behind all of it. Here is what happens when you press it.
How it works
What happens between ⌘G and the light, and why none of it needs your whole page, your history or a guess.
Press ⌘G (Ctrl+G on Windows) and say it the way you would ask a colleague: “where do I stop paying”. The page says “cancel the project from the billing page”. ⌘F would never connect the two.
A small model that ships inside the extension reads every passage, button and field (147 here, up to 3,000 on a long page) and keeps the eighty closest to what you asked. Nothing has left your computer yet.
Only those eighty snippets travel, written to your Sent Log before they go. Jev, a decision model by TypeSafe, answers “does this answer it?” for each with a calibrated probability. It writes nothing, keeps nothing and learns nothing from you.
The best answer shines and its key sentence shines brighter. A close call glows softly and asks “Did you mean…”; when nothing fits, it says so. Enter takes you there.
From our latest test of the extension, on pages it had never been tuned on: 50 questions asked on 10 real websites. The first answer it lit was the right one 38 times out of 50, and the right answer was in its top three for 43.
Eighty snippets leave your machine. Here is what happens to them.
Privacy
Only the background part of the extension can reach the network, and it writes each request to your Sent Log before it goes, byte for byte. This page keeps one too: it has been recording what the demos above would have sent.
Your query and short snippets of the page's text, never the whole page, and the site's hostname, never the full address.
Descriptions of the form's fields and the facts from your notes that could go into them; a line it cannot read as a fact goes as you wrote it, cut at its commas, only to a form with a box that could take it. Never a passport, NI or bank number, never what is typed in the fields; never password, payment, hidden or file fields.
Nothing, unless you switch on the optional cloud check, and then only the pasted text with every secret already replaced by [REDACTED].
Cookies, page HTML, full page text, form values, passwords, card details or your browsing history. Read the full policy.
All of it, free during early access.
Pricing
Pharos is free during early access: 50 actions a day, every feature in it, no card and no account.
$050 actions a day,
no card, no account
No card and no account. It just starts.
What counts as an action. One search, one tab hunt, one form filled, or one screen of a feed checked against your rules: one request to the model, written to your Sent Log first. Guard never counts, and neither does a private site you keep local, because nothing there leaves your machine. Always: zero data retention, never used for training.
Bring your own key. Pharos also runs on your own OpenRouter or TypeSafe key, set in its settings. Requests then go from your browser straight to that provider, never through us, and nothing counts against your day.
The extension's plan is for individuals. Pharos for your site is licensed per site.
Pharos on your docs, app or help centre. Early access.
Before you ask, the questions that come up first.
Questions
No to both. What Pharos sends goes through our service to the decision model, Jev by TypeSafe, running on Cloudflare Workers AI. The provider does not train on inputs, requests through our route are zero data retention, and our servers keep counts, never content: how many actions, how many tokens, which day. Our gateway keeps no log of request or response bodies and stores no answers.
On private sites (mail, documents, chat, banks and work tools) Pharos asks once per site, Local or Cloud, and nothing is sent unless you choose Cloud. Everything that does leave is in your Sent Log first, byte for byte. Read the full policy.
Because two things only recently got good enough, and Pharos needs both. A small embedding model now fits inside the extension (about 8 MB, nothing to download) and reads the whole page on your device, so only the eighty or so snippets closest to your question ever leave it.
Then a decision model, Jev by TypeSafe, answers yes-or-no and which-one questions about that shortlist with calibrated probabilities, in about half a second and for a fraction of a penny. A chat model writes prose and sounds just as sure when it is guessing; a decision model says how sure it is, and that number is what lights the page. Until recently the same trick needed a large model, several seconds and a confident guess.
Try it on the next page that makes you give up. ⌘F matches exact words, and pages rarely use yours: the refund rule says “Delay Repay”, the cancel link says “End your contract”. Pharos matches what you mean, so the first time it lights an answer that shares none of your words, it feels like a trick.
From our latest test of the extension, on pages it had never been tuned on: 50 questions asked on 10 real websites. The first answer it lit was the right one 38 times out of 50, and the right answer was in its top three for 43. It never pretends: when two places fit, it says "Did you mean…" and shows you both with their odds. Then it keeps paying its way: one key fills a form from whatever you copied, and a key from 100+ services gets caught before it lands in the wrong chat.
Your query and a few snippets of the page for Find; field descriptions and your note lines for Fill; nothing for Guard, unless you switch on its optional check. Every request is in the Sent Log before it is sent. The privacy section above is the whole list.
Literal matches appear as you type. The model's answer lights the page in about half a second, because it answers with probabilities and writes no prose. Guard runs locally and answers in well under a millisecond.
Any page you can read: documentation, dashboards, forms, feeds, webmail, and every tab you have open. You can switch it off per site in one click, or pause it everywhere.
Guard is local by default, with an optional check you switch on. On private sites such as mail, documents and work tools, Find asks once per site: Local runs on your device only and sends nothing, Cloud ranks by meaning. Elsewhere the meaning-aware ranking uses our hosted service, with zero data retention, or your own OpenRouter or TypeSafe key if you bring one.
Pharos stops and tells you when your actions come back: the next day. Rules pause with a line on the feed, so nothing slips through while you think they are on. Guard carries on exactly as before, because it never counts, and so does Find on private sites you keep local, which runs on your device.
Right after Chrome's own ⌘F, yes: the next ⌘G within a few seconds is find-next, as before. The rest of the time ⌘G opens Pharos, and you can choose another key in settings.
Say what you want. We write once, when it is in the Chrome Web Store.
