Home

navette — the browser for agents

One tiny Rust binary driving the WebView your OS already ships — no Chromium, no download, no RAM bonfire.

Navette is French for shuttle — the small vessel that carries your agent from page to page. Always fueled (the engine ships with your OS), light enough to ignore, and it skips the human web's garbage so your agent doesn't have to.

Current release: v1.9.0 (2026-10-08) · CI green on macOS (WKWebView), Windows (WebView2) and Linux (WebKitGTK) · slabbdev/navette · site: slabbdev.github.io/navette

$ ls -lh target/release/navette
-rwxr-xr-x  1 user  staff   725K  navette

725 KB installed on macOS. Playwright ships 218 MB. Lightpanda ships 96 MB (and cannot screenshot). Faster than Playwright + Chromium on every metric measured — install, cold start, navigate→read (8–11 ms), act (1–2 ms), peak RAM (79 MB vs 599 MB), 100-page crawl, real-web success rate (95–100% vs 85%). All numbers reproducible: see Benchmarks.

What it is

navette is a single Rust binary with three modes:

Mode	What it gives you
navette serve	an HTTP JSON API on 127.0.0.1 — curl-able from any language or shell
navette mcp	an MCP stdio server for agent hosts (19 tools) — auto-starts serve if nothing is listening
navette install-daemon	a resident LaunchAgent / systemd unit / scheduled task — cold start an agent feels drops to 24–37 ms

It drives the system WebView — WKWebView on macOS, WebView2 (Chromium) on Windows, WebKitGTK on Linux. Zero engine download, full rendering, real JS, real sessions. It is not a scraper but an actor: it opens sessions, fills forms, clicks through flows, verifies with screenshots, and keeps state across steps.

Start here

If you want to…	Read
install and run it in 60 seconds	Getting Started
call it from a shell / script	CLI Reference and HTTP API Reference
plug it into an agent host	MCP Integration
log into sites without handling passwords	Secrets and Logins
run it in a container	Docker
automate it from n8n workflows	n8n Integration
drive it from chat (@navette)	Chat Plugin
teach it a new website	Site Recipes
understand what runs where	Architecture
know exactly what it guarantees	Security
see the measured claims	Benchmarks
see it used for real	Demos
know the honest limits	Roadmap and Known Gaps
catch up on versions	Releases
build / CI / debug crashes	Development
the "why" of it all	FAQ and Philosophy

The one-minute tour

brew install slabbdev/tap/navette   # macOS arm64
navette serve --port 8765           # HTTP API on loopback (window 2, or --daemon via install-daemon)

curl -s -X POST http://127.0.0.1:8765/navigate -H 'Content-Type: application/json' \
     -d '{"url":"https://example.com","with_content":true,"format":"markdown"}'
# → {"ok":true,"url":"…","title":"…","content":"…"}

Or through MCP — the host gets 19 tools: navigate, read, screenshot (returned as image content — the agent sees the page), click, hover, type, key, evaluate, wait, scroll, upload, viewport, sessions, session_close, session_show, session_hide, login, state_export, state_import.

Design pillars

1. The engine is the OS. System-first, embedded fallback (WebView2 Fixed Version, WebKitGTK via apt, WPE). We will not vendor a browser the way Playwright does — that trade is the thing navette exists to refuse.
2. Agent-first surface. 8 primitives (navigate, read, screenshot, click, type, evaluate, wait, sessions), no DevTools baggage, no CDP.
3. Human-moments by design. One-time human steps (OAuth, 2FA, captcha) happen on-screen via session_show — then the window parks away and the state stays. Captchas are never solved by the machine.
4. Secrets never cross the wire. Logins live in the OS keychain, bound to origins; the agent fills a form with POST /login and never sees the password. There is no read-back API at all.
5. Measured, not asserted. Every public claim is backed by a reproducible bench (Benchmarks), and known gaps are stated plainly (Roadmap and Known Gaps).

---

MIT licensed. Built on the shoulders of tinyjs wrap (tinyjsapp). Mascot: a grumpy shuttle. It does the nav. It's literally its job.