---
title: ~1 MB to orbit: your agent's browser ships with your OS
published: true
description: A single Rust binary that drives the WebView your OS already ships — 0.6–1.2 MB, no Chromium, MCP-native. Benchmarked against Playwright and Lightpanda.
tags: showdev, rust, ai, webdev
---

Houston, we deleted Chromium.

Every AI agent that touches the web today drags the same luggage: headless Chromium.

Playwright downloads ~400 MB of browser on every machine your agent touches. Cloud browser APIs (Browserbase, Steel, Hyperbrowser) rent you Chrome by the subscription — and your pages leave your machine. And all of this to do what an agent actually needs: open a page, read it, take a screenshot, click a button, fill a form.

Meanwhile, your computer already ships a full browser engine. macOS has WebKit. Windows has WebView2 — which is Chromium, preinstalled. Linux has WebKitGTK.

So I built navette (French for shuttle): a single Rust binary that drives the WebView your OS already ships, and speaks MCP — the protocol agents already speak.

≤ 1.2 MB installed. No Chromium. No download.

One binary, three modes:

navette serve --port 8765     # HTTP API on loopback
navette mcp                   # MCP stdio server for agent hosts
navette install-daemon        # resident: warm from login

Sixteen MCP tools as of v1.4. The core loop: navigate, read, screenshot, click, type, evaluate, wait, sessions, session_close — since joined by scroll, file upload, viewport control, in-page dialog handling, and session state export/import. The screenshot comes back as MCP image content — your agent literally sees the page. Full tool list in the repo.

Under the hood on macOS: a WKWebView per session living in a ghost window (real pixels, parked off-screen so it renders but nobody sees it). Navigation completion is event-driven — I implemented WKNavigationDelegate in Rust with raw objc2 message sends, and the content extraction runs inside the didFinish callback, so a warm navigate + full markdown read costs 8 ms.

Same machine (M1, 8 GB), same corpus (100 local pages + 20 real URLs), reproducible with one command (python3 bench/navbench.py). macOS numbers; cross-engine medians and p95s run in CI (see the repo):

Two honest disclosures, because benchmark posts die without them:

1. Lightpanda is the fastest page-reader (3.3 ms through a zero-bias raw-CDP probe — I measured it through a fair minimal client, not just through Playwright). It parses a partial DOM and cannot render, screenshot, or act. If your agent only reads static pages, use it — or plain fetch + readability.
2. The two rows Lightpanda wins (fresh boot, peak RAM) are the price of rendering. Three WebKit helper processes and a backing store are what buy screenshots and full fidelity. That tax is the product.

The claim I'll defend: nobody else combines ≤ 1.2 MB + 1 ms actions + a 0.9 s hundred-page crawl + full WebKit-class rendering + a resident 24 ms mode.

I pointed the resident daemon at a login form and a research loop and let an agent drive through MCP only:

• Act 1 — authenticate: type credentials, click, verify the session ("You logged into a secure area!"), screenshot the authenticated page.
• Act 2 — research: DuckDuckGo → extract results → open one (it was dead — moved on) → read the article → screenshot.

Full walkthrough with the verbatim tool calls and screenshots is in demo/JOURNEY.md.

brew install slabbdev/navette/navette     # macOS arm64
cargo install navette-browser             # any platform
docker run -i --rm ghcr.io/slabbdev/navette navette mcp   # container, stdio MCP

# or from source:
cargo build --release
./target/release/navette serve --port 8765

# or let your agent host own the lifecycle:
./target/release/navette mcp

Register it in any MCP client (ZCode example):

{ "mcp": { "servers": { "navette": {
    "command": "/path/to/navette", "args": ["mcp"] } } } }

Then just talk to your agent: "open this dashboard, check if the deploy banner is there, screenshot it for me."

• Three engines, one fully hardened. Windows (WebView2) and Linux (WebKitGTK) reached platform parity in v1.4 — screenshots, cookie state, viewport, resident daemon, CI-verified on all three OSes — but macOS remains the benchmark-carrying backend. The engine strategy is written down: system-first, official embed (WebView2 Fixed Version) or distro packages as fallback. Never vendor a browser the way Playwright does.
• The automation long tail is still long. Still missing, stated plainly: OS-level keyboard/mouse input, request/response network interception, OS file-dialog automation — Playwright has twenty years of API surface. The core agent loop (navigate, read, see, act, wait, sessions, state) is complete; the tail is the roadmap.
• Fresh-process cold start can't beat a blind engine. A full WebKit spawns three helper processes; Lightpanda spawns none because it renders nothing. That's why navette ships a resident mode: warm from login, 24 ms forever.

Headless Chromium is the Postgres of web automation. navette is trying to be the embedded SQLite — the browser agents bring with them, for the fastest-growing half of the agent world: agents that run locally, on machines that already have an engine.

Apple shipped a Safari MCP server for coding agents this year. The thesis is being validated from above. navette is it from below: tiny, open, cross-platform.

• v1.2.0 — Oct 2: Windows (WebView2) and Linux (WebKitGTK) backends aboard; state_export / state_import — cookies and storage out of the browser and back in, the Playwright storageState equivalent. 11 tools. Release
• v1.2.1 — Oct 3: CI green on all three OSes (same navigate→read smoke). Getting there surfaced five real root causes — fold-result routing, URL-matched completion, a missing D-Bus session bus in headless Linux, WebView2's file:// stall — each fixed and proven with a breadcrumb or backtrace. Boot-time pre-warm parity: first navigate on a cold CI VM went 38 s → 83 ms. Release
• v1.3.0 & v1.4.0 — Oct 4: full platform parity (native screenshots per engine, cookie round-trip, viewport control, resident daemon — CI verifies a valid PNG and a cookie round-trip on all three OSes); JS dialogs auto-handled in-page; the agent loop completed with file upload (in-memory DataTransfer, no OS dialog) and scroll — 16 tools; fixed an eval-wrapper bug where every /evaluate and /read outside the fold path returned undefined on Windows/Linux (CI now tests it); one-command install cargo install navette-browser. Release
• v1.5.0 — Oct 5: the idle watchdog — serve --idle-release 15 drops WebKit sessions after 15 minutes without requests while the daemon stays resident; the next request re-warms on demand (38 s cold → 83 ms warm on the CI VM). Plus the tollbooth bench harness: ~200 real URLs × 5 categories, measuring robots.txt AI clauses, llms.txt adoption and the agent path itself — the walled-web dataset is committed to the repo (its own post soon). Release
• v1.6.0 — Oct 7: the operator release — --token (Bearer auth on the HTTP API, MCP hosts attach with NAVETTE_TOKEN), --proxy (HTTP CONNECT / SOCKS5), --user-agent (per-serve override); linux-arm64 joins the release assets; cross-engine medians (warm navigate+read 5.7–23.4 ms on all three OSes) landed in BENCHMARKS.md. Release
• v1.7.0 — Oct 7: native keyboard input — POST /key attempts real OS-level events first (SendInput on Windows, XTEST on Linux, CGEvent research parked on macOS) and falls back automatically to the synthetic dispatch, always reporting which mode ran. 18 MCP tools. Release
• v1.8.0 — Oct 8: visible session windows (session_show/session_hide — a ghost session comes on-screen with its page state intact for one-time human logins, then parks off-screen again) and pointer-event clicks (pointerover/pointerdown/pointerup before the mouse events — Radix/headless-UI components finally open), the macOS /wait boolean-stringification bug fixed, the backend contract became the compile-enforced webviewkit kit, and the tollbooth bench went round 2 — x402/paywall signals, dataset CC BY 4.0. Release
• v1.8.1 — Oct 8: cross-session cookie isolation is now a build-breaking CI assertion on every engine — named WebView2 profiles + InPrivate on Windows (InPrivate alone shares one profile across controllers, measured), non-persistent stores on macOS, isolated contexts on Linux — plus loopback hardening: constant-time token comparison, a DNS-rebinding guard, credential-redacted proxy logs. Security posture documented in docs/SECURITY.md. Release
• Sizing, honestly: the title's 594 KB was v1.0. macOS arm64 is 658 KB today, Windows x64 / Linux x64 ~1.2 MB — still ~180× smaller than Playwright's browser download.

Repo, benchmarks and the reproducible harness: https://github.com/slabbdev/navette

If you run agents locally, I'd genuinely love your feedback — especially the failure cases.