Here's a bet worth making.
Your browser has watched you read for years. Every article, every review, every half-finished bit of research at eleven at night. And it kept a list of URLs. Titles and timestamps. It threw the actual pages away — everything you were there for — and left you with a history panel you've probably never successfully found anything in.
That's a strange thing for a piece of software whose entire job is reading the web. It's a stranger thing for Chervil, which composes pages for you and then, until today, forgot them just as thoroughly.
0.20.0 fixes that. Chervil now keeps a private index of your own reading — the pages it makes for you, and (if you want) the text of the sites you visit. It lives on your machine. And Sprig can answer from it.
"What was that laptop review I read last month?"
Ask that. Out loud, in the way you'd ask a person.
Chervil goes to your own history — not the web — finds the pages you actually read, and answers from them. It tells you which page, which site, and when you read it, with a link straight back. No search engine involved. No tokens spent. Works on a plane.
This is the thing a search engine structurally cannot do. Google never saw the page you read. It doesn't know you read it. Your browser watched the whole thing and — until now — chose not to remember.
And when Chervil can't find it, it says so. It won't quietly hand you a web answer dressed up as your own page, because the worst possible outcome here is you believing you read something you never did.
Pull it together
Re-finding is the small version. Here's the bigger one:
Pull together what I've read about mirrorless cameras
Sprig gathers everything on the subject from your reading and synthesizes across it — the through-line, where your sources contradict each other, and, most usefully, what you haven't looked at yet. Everything is attributed to the page it came from, so you can get back to any of it.
That's the research you were going to do yourself, at eleven at night, across twelve tabs you'd already half-forgotten.
It can also notice
If you turn it on, Chervil watches for a subject that keeps coming up across several different sites — the signature of actually researching something, as opposed to just reading one site a lot. When it sees one, a quiet card appears on the new-tab screen:
You've been reading about mirrorless — 5 pages across 4 sites. Want me to pull it together?
Two things about that card, both deliberate.
It offers. It never writes the page on its own. Nothing is composed and nothing is spent until you click. An ambient feature that guesses wrong occasionally is fine if a wrong guess costs you one dismissible card. It is not fine if it costs you tokens and a surprise essay about something you didn't care about.
It never interrupts. No popup, no sound, no notification. It waits on the new-tab screen until you happen to look. You didn't ask for it, so it doesn't get to demand your attention. "Not interested" means never again for that subject.
How eager it is, is a dial you control. It's off by default.
The privacy part, which is the whole thing
Keeping the text of everything you read is the most sensitive thing this app has ever done. It's also the sharpest argument for it — but only if it's airtight. So:
- It's a file on your computer. Never synced, never uploaded. Not to me, not to anyone.
- No AI sees a page because you read it. Nothing is sent anywhere when you read. Nothing is summarized in the cloud. The text only reaches your AI provider if you ask a question it answers — exactly like attaching a file, and only then.
- It's off until you turn it on. Composed pages — the ones you knowingly made — are indexed either way. The sites you read are opt-in, in Settings → Security → Your Web.
- Private tabs are never included. Neither is anything behind a login or a paywall.
- You can skip any site you name. Your bank, your health portal, whatever you like.
- You can see what's kept, and forget the last hour, today, one site, or everything. Those deletes are real — the rows are removed and the file compacted, not hidden from a list.
There's a design decision underneath all of that worth naming, because it's why the promises are cheap for me to keep. Chervil doesn't run a model over pages when you read them. It stores the text and lets a local, decades-old full-text search rank it, and only asks the AI anything at the moment you ask a question. Which means reading costs nothing, works offline, needs no API key — and, decisively, nothing leaves your machine. A "summarize every page as you go" design would have shipped everything you read to a provider. That alone disqualified it.
Ask it twice, get it instantly
A smaller thing you'll feel every day. Ask something you've already asked, and Chervil serves the page it already made — labelled, plainly:
From your pages — you asked this 3d ago. Nothing was searched or composed.
…with ↻ Answer it fresh right there. Never silently: a cached answer you can't identify as cached is just a lie with better latency. And anything whose answer moves — news, prices, "today", "latest" — is never served from cache, no matter how recent.
The top bar looks like a browser now
Chervil had twenty buttons in its toolbar. Chrome has four and a menu. Turns out no amount of rounded corners makes a twenty-icon toolbar look like a browser.
So: five buttons, and the rest moved into the ⋯ menu. Nothing was removed — snip, translate, read-aloud and the others are all one click away, and you can give any button a permanent seat back in Settings. The address bar is a proper pill with the page's status shown inside it. The active tab now flows into the toolbar instead of hovering above a dividing line.
And Settings is a page now, not a dialog — it opens in its own tab, with a nav rail down the side, and it remembers which section you were on.
Also in 0.20.0
- Attach a photo and ask Sprig to change it, and he changes the photo. Previously "replace the book with a calculator" would compose a page with broken images — or, memorably, a hand-drawn cartoon captioned "Image Successfully Edited". Sprig now knows what he can't do and says so, and an edit request opens the real image editor.
- Chervil opens faster, and composed pages appear instantly when you switch back to them instead of being rebuilt from scratch every time.
- Your pages moved into a real database, so the session file stopped being rewritten every time you type.
Chervil is free and open source. Grab 0.20.0, or read the design if you'd like to see exactly what it does and doesn't keep.
The bet, stated plainly: the browser is the only piece of software that sees everything you read. It should be the one that helps you remember it — and it should never be the one that tells anyone else.