One Agent - One Browser

This is an experiment to see if an agent using LLMs, could build a functional browser by itself with minimal guidance, and without using any 3rd party libraries.

Ideal Result / Goals

• A binary that can render a .html file to the display

Running

Platform:

• Linux/X11 (requires an X server and $DISPLAY).
• Windows 10 (1703+) or Windows 11
• macOS 11+ (double-check)

System dependencies

This project uses system libraries/frameworks via FFI.

• Linux/X11: X11/Xft/Cairo, librsvg, libcurl, libpng, libjpeg-turbo, libwebp.
• Windows 10/11: WinHTTP, WIC (PNG/JPEG/WebP), Direct2D/DirectWrite. If WebP decode fails, install Microsoft "WebP Image Extensions".
• macOS: system frameworks (CoreGraphics/CoreText/ImageIO/QuickLook).

Arch Linux (Wayland via XWayland):

sudo pacman -S --needed xorg-xwayland libx11 libxft cairo librsvg curl libpng libjpeg-turbo libwebp

Ubuntu (Wayland via XWayland):

sudo apt-get update
sudo apt-get install -y xwayland libx11-dev libxft-dev libcairo2-dev librsvg2-dev libglib2.0-dev libcurl4-openssl-dev libpng-dev libjpeg-turbo8-dev libturbojpeg0-dev libwebp-dev

RHEL (Wayland via XWayland):

sudo dnf install -y xorg-x11-server-Xwayland libX11 libXft cairo librsvg2 libcurl libpng libjpeg-turbo libwebp

If you run an Xorg session (not Wayland), install an Xorg server package instead of XWayland (xorg-server / xorg / xorg-x11-server-Xorg).

# Built-in "Hello World"
cargo run

# Render a local file
cargo run -- test-file.html

# Render a URL
cargo run -- https://example.com

# Save a PNG screenshot and exit once the page is ready
cargo run -- test-file.html --screenshot out.png

# Headless mode (Linux/X11: still requires an X server, e.g. via xvfb-run)
cargo run -- --headless test-file.html --screenshot out.png

Arguments

• <target> (optional): path to an HTML file, or an http(s)://... URL.
• --screenshot <path> / --screenshot=<path>: write a PNG screenshot and exit.
• --headless: don't map a window; useful for automation/tests.
• --width <px> / --width=<px>: initial viewport width in CSS pixels (default: 1024).
• --height <px> / --height=<px>: initial viewport height in CSS pixels (default: 768).
• OAB_SCALE (env): override the DPI scale factor (e.g. 1.25 or 125%).

Tests

cargo test

# If you don't have an X server (Linux CI/headless), use Xvfb:
xvfb-run -a cargo test

Render regression tests compare screenshots to per-platform baseline PNGs in tests/cases/.

• OAB_RENDER_TEST_MIN_SIMILARITY (env): minimum required similarity ratio (default: 0.95; set 1.0 for exact match).