Chervil — the agentic, conversational web browser

⚠️ Alpha / building in public. Chervil is young and moves fast. Many
features are implemented but not yet end-to-end verified on every platform.
Treat this as a preview, not a stable release. A Windows installer is available
on the Releases page (the setup
wizard can set your keys, default provider, profile, and run-at-startup); you can
also run from source (below).

🌐 Website: getchervil.com · 📣 Announcement: Coming Soon — Chervil: The Web Comes to You

Chervil replaces the browse-and-skim loop with a conversation. Instead of ten
blue links, you talk to Sprig — an assistant that searches the live web, then
composes a single, self-contained, image-rich page built for your exact question.
When you actually want a real site (email, bank, a web app), Sprig opens it live.
Synthesized pages when synthesis is better; the real web when the real web is the
point.

(Chervil is "French parsley" — which is exactly why Sprig, our leafy guide, feels
right at home.)

Demo

🎬 A short screen capture of Sprig composing a page from a question is on the
way. (Building in public — it'll land right here.)

What it does

Compose pages, not link lists — grounded in live web search, rendered in a sandbox.
Living, interactive pages — composed pages can call back to Sprig at runtime to fetch fresh data (mini-apps, not printouts).
Deep Dive — a two-phase research pipeline that produces a long, cited report with disinformation vetting.
Trust layer — every page can show its sources and one-click Verify its own claims against live sources.
Spaces — persistent topic workspaces that synthesize across everything you've gathered.
Living pages — schedule a page to re-ground itself and notify you when it changes.
Agentic web actions — Sprig can operate real sites for you, with hard safety gates (see SECURITY.md).
Quick-ask — a global hotkey opens a floating ask bar; Chervil lives in the tray, always a keystroke away.
"Hey Sprig" listening — opt-in hands-free wake phrase that pops the quick-ask bar and captures your spoken request. Wake-word detection is fully on-device (Picovoice Porcupine) — audio never leaves your machine until you speak a command. Needs a free Picovoice key; the literal "Hey Sprig" phrase uses a custom keyword file (built-in words work out of the box).
Scheduled agents — run prompts on a daily/weekly/interval schedule in the background, optionally "as" a chosen agent.
Agent files — import a persona (instructions, model, tools, starters) to make Sprig a specialist. Starters in /agents.
Video summaries — summarize YouTube videos from captions, or watched natively on Gemini.
Export anywhere — turn any page into PDF, PowerPoint, Word, Excel, an image, or an animated GIF.
Bring your own AI — pluggable providers: Claude, Grok, Gemini, OpenAI, Azure AI Foundry, and local Ollama.

Run from source

git clone <this-repo>
cd <repo>
npm install
npm start          # alias: npm run dev

You'll need an API key for at least one provider (or a local Ollama
install for free, offline use). Add keys in Settings ⚙ → Provider — they're
encrypted at rest on your machine and never round-tripped through the UI.
Environment variables also work (a .env is supported):

Variable	Default	Notes
`CHERVIL_PROVIDER`	`claude`	Model backend: `claude`, `grok`, `gemini`, `openai`, `azure`, `ollama`.
`CHERVIL_MODEL`	`claude-sonnet-4-6`	Use `claude-opus-4-8` for maximum quality at higher latency.
`CHERVIL_EFFORT`	`low`	How hard the model deliberates (`low`/`medium`/`high`).
`ANTHROPIC_API_KEY`	—	Claude — the only provider with the full Deep Dive + web grounding (search and fetch).

Provider notes: Claude has the richest experience (live grounding + two-phase
Deep Dive). Grok, Gemini, and OpenAI support live grounding. Azure and Ollama
are compose-only (no live web search).

Env vars: CHERVIL_* is the primary prefix. The older PARSLEE_* (and
legacy PINGCHAT_*) names are still read as fallbacks, so existing .env
files keep working.

Status & roadmap

This is an early public build. See docs/pre-launch-hardening.md
for the known verification gaps we're closing before a signed, packaged release.
The direction: deeper multi-step agents, Spaces that hold your files as permanent
sources, richer computed pages, and ever-stronger trust tooling.

Tech

Electron desktop app (bundles its own engine — no installed browser needed). A
main process holds keys and talks to providers; a sandboxed renderer is the UI;
the model layer is fully pluggable. See CONTRIBUTING.md for
architecture and how to add a provider.

chervil