Type what you want.
Get the right sites.

Every other AI browser answers for you. Jasb doesn't. It reads what you meant, finds the handful of pages that actually serve it, and gets out of the way — so you end up on the web instead of reading a summary of it.

Try it right here Get the browser

No account No ads, ever Open source, MIT History stays on your device

Try:

Three ways in

One engine. Pick the shell that fits your week.

The ranking logic is one package, so every one of these returns the identical grid for the identical query. What differs is how much of the browser you want us to be.

How it works

The order of these steps is the whole business model.

Most searches never reach a paid API at all. That is not an optimisation bolted on afterwards — it is why a browser like this can exist without ads.

  1. 01 Recognise a destination A URL, a bare domain, localhost:3000, or a bang like !w goes straight there. No model runs. free
  2. 02 Check the navigation index “figma sign in”, “irs login” — single-destination requests resolve from a local brand map in about ten microseconds. free
  3. 03 Check the cache Exact match first, then a semantic near-match that a decision model has to confirm before anyone gets someone else's results. free
  4. 04 Read the intent What kind of thing is this, and which sources suit it? One call, two typed questions, answered with calibrated probabilities rather than prose.
  5. 05 Ask several indexes at once Brave for the open web, Wikipedia for reference, Marginalia for the small web, your own SearxNG if you run one. A slow source is dropped, not waited on.
  6. 06 Measure quality Tracker load, page speed, paywalls — plus anything this device has learned from pages you actually opened. Your block and pin rules apply last, locally.
  7. 07 Judge every candidate Relevance, spam probability and a label, for all twenty candidates in parallel. Cheap enough to do on every single query.
  8. 08 Show six — or fewer When the judgements come back genuinely unsure, you get three cards and a note saying so, instead of six with three of them padding.
By design

The list of things it won't do is the product.

Every item here is something a competitor ships and we deliberately don't. None of them is a roadmap gap.

The unfair advantage

A browser can measure what a crawler has to guess.

Ranking by page quality normally means crawling the web and estimating. We are already in the browser: the moment you open a page we know how many third-party hosts it contacted, how long it took, whether it hid behind a wall, and whether you bounced within three seconds.

That evidence lands in a database on your own machine and improves your ranking first. It is never uploaded. Sharing it in aggregate would need your explicit permission and a k-anonymity threshold — and it isn't built yet, so it isn't claimed.

Pricing

One low price, or none at all if you bring your own keys.

An uncached search costs us about half a cent, almost all of it the search API rather than the model. The pricing just reflects that.

Paid plans are not open yet. When they are, payments run through Paddle as merchant of record, so they handle tax and invoicing and we never see a card number.

Questions

The ones worth asking.

Isn't this just a search engine with extra steps?

It's a browser where the address bar understands sentences. A search engine gives you a page of ten blue links with ads on top and asks you to evaluate them. Jasb gives you six sites, each labelled with what kind of page it is, ranked partly by measurements taken in your own browser. And when you type a URL, nothing runs at all.

Why would I pay for this when Google is free?

Google isn't free — it's paid for by the ads at the top of every result and by what it learns about you. If that trade suits you, it's a genuinely good deal and you should keep taking it. Jasb is for people who would rather pay five dollars and have neither. And if you'd rather pay nothing, bring your own API keys: that tier is free forever and we never see the traffic.

What do you actually store?

On our server: a hash of the query mapped to the resulting list of sites, and only once at least three separate devices have asked for that query. No IP addresses, no query text, no account. On your device: history, favourites, your block and pin rules, and the page measurements — all in one file you can erase with one button. The privacy policy spells it out.

What happens if your servers go away?

The desktop app runs the whole engine locally. Point it at your own keys — or at a self-hosted SearxNG and an open decision model — and it never needs us. The source is MIT-licensed, so that stays true whatever happens to the company.

Which AI model is this, and can I run it myself?

Mostly not a chat model at all. The work is selection and scoring, which is what “System One” decision models do — they return calibrated probabilities instead of text. We use the open Apache-2.0 ones (Kev, CLM, Von, Decider), all of which you can run on your own hardware, and fall back to an ordinary LLM through your own key when you'd rather. Jev, the closed hosted one that defined the format, is supported but is not the default.

Is it any good? Honestly.

Try the box at the top and judge for yourself — that is the real engine, not a recording. There's also a blind comparison tool in the repository: a hundred real queries, our six results against a competitor's first six, sides shuffled and unlabelled until you've graded every one. We built it to find out whether this was worth shipping, and we'd rather you ran it than took our word.

The web is still out there. Go and read it.

Try it right here Read the source