Chervil gives you two ways to work with Sprig. They share the same conversation and the same live-web access — the difference is the output.
Compose mode
This is Chervil's default. You ask for something and Sprig composes a complete, self-contained page to answer it — with real typography, tables, cards, and interactivity where it helps. Composed pages can:
- Search the live web and cite sources inline — when Sprig uses the web, the specific claims it draws from a source get a small numbered citation chip, with a matching numbered Sources list at the bottom.
- Refine in place — the page's action bar has a follow-up box: type "make it shorter," "focus on the budget options," or "add a section on X" and Sprig updates the page you're looking at rather than starting over.
- Include live applets — small interactive tools Sprig builds into the page.
Every composed page has a Sources panel showing the live pages Sprig used — now with a freshness summary (e.g. "Freshest source 2 days ago · 4 of 6 dated") and a recency badge on each source, so you can see how current the answer is at a glance. If a page answered from the model's own knowledge instead, the panel says so — a good cue to run Verify.
Compare things side by side
Ask Chervil to compare two (or more) things — iPhone 16 vs Pixel 9, two frameworks, two
towns — and instead of prose it builds a sourced comparison table. Trigger it by typing a
plain "X vs Y", pressing the ⚖️ button by the composer, or with /compare ….
Sprig researches the live web and lays out your options as columns and the decision factors as rows, flags the winner of each row, and adds a bottom-line verdict plus a "best for…" guide that matches each option to the person it suits. Every cell is grounded in a real source. Click a column header to focus it, or toggle Differences only to hide the rows where everything's the same.
Watch a page for changes
Chervil can keep an eye on a page for you and speak up when something changes. On any live site, just ask — "watch this page and tell me when it's back in stock", "let me know when the price drops below $500", "ping me when the docs update". You can also add one on 🗓 Sprig's Desk by giving a standing goal a page URL instead of a topic.
Sprig checks the page in the background on a schedule and sends you a notification the moment your condition is met (or, if you didn't give one, whenever the page's key detail changes) — clicking the alert opens the page. Watching a page is a standing goal now, so it lives on the Desk alongside everything else Sprig keeps an eye on — see Sprig's Desk for the full picture. It runs while Chervil sits in the tray, and syncs across your machines.
Living pages: what changed, and watching a number
Any composed page can keep itself current. Pick an interval from the ⏱ Live dropdown on the page's toolbar and Sprig re-runs its question on that schedule and updates the page in place.
Because the page is genuinely re-composed each time, its wording changes almost every refresh — so instead of just telling you that something changed, Chervil compares the figures:
starting price $899 → $849 · launch date Sept 12 → Sept 19 · warranty cover 24 months (new)
You'll see a change count next to the live indicator, and the list in the Sources panel. If only the wording moved, Chervil says so — "reworded, no figures changed" — which saves you re-reading it.
Tell me when a number crosses
On a page that's keeping itself current, just ask:
- "Tell me when the price drops below $1,500."
- "Let me know if battery life goes above 30 hours."
- "Alert me if the launch date changes."
Chervil watches that figure and notifies you when it crosses. The check runs on the refresh the page was already doing, so it costs nothing extra, and it fires when the number enters the range you asked about rather than every time it moves around inside it. Everything you're watching is listed — and cancellable — in the Sources panel.
If the page isn't live yet, asking this turns it on (checking every 30 minutes).
This is the sibling of Watch a page for changes above. That one polls a live website; this one watches a figure on a page Sprig composed from many sources.
When Sprig gets something wrong
Correcting the page in front of you fixes that page. To fix it permanently, type
/correct followed by what's actually true:
/correct our office moved to Dayton in March
Sprig applies that to every page it composes from then on, so it stops repeating the mistake. Corrections are listed in Settings → You, each with its date, and you can remove any of them with Forget.
Two things to know about how they're treated:
- They beat the model's training, not today's evidence. Each correction is dated. If Sprig searches and finds that the world has moved on — a price changed, someone left a role — it goes with what it found and tells you, so an old note can't quietly outlive the truth.
- They're applied narrowly and silently. A correction about one product says nothing about the whole category, and Sprig won't announce that it was corrected — it just gets it right.
It's an explicit command on purpose. A "correction" stored because the app guessed you were correcting it would be worse than one you had to type.
Chat mode
Sometimes you just want to talk, not get a whole page back. Switch to chat mode and Sprig replies in plain conversation in the panel — quick questions, brainstorming, follow-ups.
Chat mode is fully web-aware on providers that support search (Claude, Grok, Gemini, OpenAI): Sprig decides on its own when a question needs current information and searches for it, listing the sources it used right under the reply. Ask "what's the weather tomorrow?" or "who won last night?" and you'll get a grounded answer without leaving the conversation. For evergreen questions it just answers directly.
Chat mode also sees the page you're looking at. If Sprig composed a page and you switch to chat, you can ask questions about "this page" and Sprig answers from its contents.
Replies stream in as Sprig writes them, so you can start reading immediately. The send button becomes a Stop while he's answering — and stopping keeps whatever he'd already written, rather than throwing the reply away.
On providers without web search (Azure, Ollama), chat answers from the model's knowledge and will suggest composing a page when a request needs the live web.
Keeping an answer
Chat's one weakness used to be that it forgot. A composed page files itself in your Library automatically, but a chat answer lived only in the tab it was born in — close the tab and the good one went with it.
Hover any of Sprig's replies and a ☆ appears next to the copy button. Click it and the answer is kept; click the ★ to unkeep it. Kept answers land in Library → Replies, each one filed under the question that produced it, with the answer's opening lines, when you saved it, and how many sources it used.
From there you can:
- Read it back — the Library search matches inside your saved answers, not just their titles, so you can find the one about GC pauses by searching for "GC pauses."
- Reuse — drops the answer into the composer as quoted context, so whatever you type next acts on it: "turn this into an email," "what's changed since?", "make this a checklist."
- Copy — the raw text, straight to the clipboard.
- Open it as a page — click the row and the answer becomes a real Chervil page, which means Export, Print, Read aloud, Remix and Save all work on it like anything else you've composed.
Saved replies sync across your computers the same way bookmarks do.
Choosing the chat model
Settings → Chat → Chat model picks which model answers you in chat:
- Faster model when available (the default) — uses your provider's quicker tier if it has one, otherwise your main model.
- Same as my main model — whatever you've selected under Provider.
- Any specific model your provider offers, read live from your own account.
This affects chat only. Composed pages, Deep Dive, skills and agents all still use your main model.
It's worth changing if chat feels slow, and the reason is worth knowing: a reasoning model thinks before it writes a single word, so you wait through that before anything appears on screen. Picking a non-reasoning model for chat can be the difference between a few seconds and a fraction of one. Your main model is untouched, so you keep the deeper thinking where it actually pays — building pages.
If your provider is briefly rate-limited (too many requests in a short window), Chervil waits and retries automatically rather than failing the reply. You'll see a short retrying note.
Which should I use?
- Compose when you want something to keep, share, or act on — a plan, a comparison, a dashboard, a lesson.
- Chat when you want a fast answer or to think out loud.
You can move between them in the same session — the conversation carries over.