A web browser engine built from scratch in Rust. Renders HTML, CSS, and JavaScript with software rendering via tiny-skia.
Incognidium rendering real-world pages at 1024px viewport (compared head-to-head against Firefox headless in the QA pipeline).
More Wikipedia pages (articles, science, history)
Currently a React SPA — incognidium's JS engine doesn't yet hydrate React trees, so the page shows the "enable JavaScript" fallback. Tracking progress.
# Desktop browser
cargo run --release -p incognidium-shell -- https://news.ycombinator.com
# Render a page to PNG
cargo run --release --bin render_to_png -- https://lite.cnn.com /tmp/output.png
# Render with text extraction
cargo run --release --bin render_to_png -- https://lite.cnn.com /tmp/output.png --text /tmp/output.txt
# Crawl sites for training data
cargo run --release --bin incognidium-crawl -- --sites news# Build .deb package (Linux)
cargo install cargo-deb
cargo deb -p incognidium-shell
sudo dpkg -i target/debian/incognidium_*.deb
# Then run from anywhere
incognidium https://en.wikipedia.orgcrates/
incognidium-dom/ # DOM tree (arena-based)
incognidium-html/ # HTML5 parser (html5ever)
incognidium-css/ # CSS parser (cssparser) + selector matching
incognidium-style/ # Style resolution, cascade, inheritance
incognidium-layout/ # Box layout: block, inline, flex, table-as-flex
incognidium-paint/ # Pixel rendering (tiny-skia) + TTF text
incognidium-net/ # HTTP fetching (reqwest + rustls)
incognidium-shell/ # Browser shell, JS engine (Boa), windowing
incognidium-devtools/ # MCP devtools bridge
Tested against Firefox headless on 20 sites:
| Grade | Sites |
|---|---|
| GREAT (diff < 25) | Hacker News, CNN Lite, NPR, Lobsters, Dan Luu |
| OK (diff 25-50) | Wikipedia, arxiv, Weather.gov, AP News, NYTimes |
| Rendering | TechCrunch, BBC, Slashdot, Nature, Ars Technica |
- External stylesheets, inline styles,
<style>blocks @mediaqueries with viewport-aware min/max-width filteringflexshorthand,flex-direction,flex-grow/shrink/basis- Multi-value
margin/paddingshorthands (1-4 values) bordershorthand with width + style + colorbox-sizing: border-box!important, specificity-based cascade- Pseudo-class filtering (skip
:visited/:hover/:focus) text-align,text-transform,text-decorationopacity,visibility,overflow,white-spaceposition: fixedelements skipped- ~120 named CSS colors,
rgb()/rgba(), hex colors em/rem/pt/vw/vhunits
- DOM:
getElementById,querySelector,createElement,getElementsBy* - Window:
setTimeout,setInterval,requestAnimationFrame,fetch - Observers:
MutationObserver,IntersectionObserver,ResizeObserver - Storage:
localStorage,sessionStorage - 50+ DOM constructor stubs (
HTMLElement,Element,Node, etc.) getComputedStyle,matchMediawith viewport awareness- Script limits: 256KB per script, 1MB total per page
- Block flow (vertical stacking)
- Inline flow (horizontal with word-wrap)
- Flexbox (row/column, grow/shrink/basis, justify/align)
- Table-as-flex (tr = flex row, td = flex items)
max-widthcentering,text-alignfor inline content- List bullet markers (ul/ol)
- Image loading and sizing
Archive the web for training data:
# Crawl news sites
incognidium-crawl --sites news
# Crawl all 40+ sites
incognidium-crawl
# Single URL
incognidium-crawl --url https://arxiv.org/list/cs.AI/recent
# View history
incognidium-crawl --history
# Corpus stats
incognidium-crawl --statsArchive saved to ~/.incognidium/archive/ as JSONL (one record per page per crawl).
The npc_team/ directory contains an NPC team for automated testing:
# From npcsh
@renderer /render_batch iterations=3
@renderer /render_compare url="https://arxiv.org" name="arxiv"MIT