Cornea
Deterministic visual inspection for AI agents. The eyes a coding agent never had.
Cornea gives a coding agent that builds web pages a real way to see them. As a token-cheap, deterministic structural model it can reason over exactly, instead of megabytes of screenshots and raw DOM dumped into context.
An agent doesn't need a photo of a page. It needs to know is my layout broken, and how. Cornea computes an abstract visual geometry model of every element (box, position, z-order, computed styles) and derives inspection conclusions. overlap, overflow, contrast, quality. Then exposes them as native tools on three surfaces: CLI, MCP, and HTTP API.
One promise holds everything together: same input → byte-identical output. No Chromium, no sub-pixel variance, no server. A single ~1.4 MB binary.
Quick start
# Build (Rust 1.98+, edition 2024)
cargo build --release
# Inspect a file, the simplest way to use Cornea
./target/release/cornea tests/fixtures/sample-bugs.html 360
Requires only
cargo. No browser, no Node, no system dependencies. Builds fine on a phone-class device.
Or install the cornea command from npm (no Rust toolchain needed):
npm install -g optrex # provides the cornea command
cornea --help
Visual guide: what Cornea does
Take this page below. It looks fine as source. But it's hiding four layout bugs. Cornea finds every one.
<!-- tests/fixtures/sample-bugs.html (excerpt) -->
<section class="row">
<div class="card">Card 1</div><div class="card">Card 2</div><div class="card">Card 3</div>
</section>
<div class="overlap-left">Left overlap</div> <!-- position:absolute; left:20; top:20 -->
<div class="overlap-right">Right overlap</div> <!-- position:absolute; left:150; top:40 -->
<div class="overflow-bad">...</div> <!-- width:600 on a 360 viewport -->
<p class="low-contrast">Hard to read on white</p> <!-- color:#cccccc on #ffffff -->
Run it and Cornea reports the damage instantly:
$ ./target/release/cornea tests/fixtures/sample-bugs.html 360
{
"html_file": "tests/fixtures/sample-bugs.html",
"viewport_w": 360.0,
"element_count": 17,
"est_tokens": 1068, // <-- entire page read for ~1k tokens
"report": {
"total_elements": 17,
"visible_elements": 14,
"overlaps": [ ... 7 collisions ... ],
"overflows": [ ... 1 clipped ... ],
"contrast": [ ... 2 AA failures ... ],
"quality": { "score": 0.06, "label": "broken" }
}
}
Each finding is precise and actionable:
| Finding | Detail |
|---|---|
| Overlap | section.row ⇄ div.overlap-right, area 20000 px². The absolutely-positioned boxes cover the cards |
| Overflow | div.overflow-bad: right edge 600 exceeds viewport 360 → clipped |
| Contrast | black on blue: ratio 2.44:1. Fails WCAG AA (needs 4.5) |
| Contrast | #cccccc on white: ratio 1.61:1. Fails WCAG AA |
That is the value: hundreds of tokens, not hundreds of kilobytes, and a deterministic answer the agent can act on and re-verify.
Visual guide: the three surfaces
Cornea is one engine, three doors. All three returns identical inspection JSON because they funnel through a single shared dispatch.
1. CLI. Inspect a file
cornea <file.html | http(s)://url> [viewport_width] [viewport_height] [--js]
$ cornea page.html 360
{
"html_file": "page.html",
"viewport_w": 360.0,
"element_count": 42,
"json_bytes": 8124,
"est_tokens": 2193,
"report": { "total_elements": 42, "visible_elements": 38, "overlaps": [], "overflows": [], "contrast": [], "quality": { "score": 1.0, "label": "good" } }
}
2. MCP. Native agent tools over stdio
cornea --serve
An agent calls layout.* tools directly; it passes the page source, or a
live URL (see URL capture below), per call:
→ {"method":"tools/call","params":{"name":"layout.overlaps",
"arguments":{"html":"<div style=\"position:absolute;left:20;top:20;width:200;height:100\">A</div>..."
,"width":360}}}
← {"id":1,"result":{"content":[{"text":"[{\"a_sel\":\"...div \u21c4 ...div\",\"area\":20000}]"}]}}
| Tool | Returns |
|---|---|
layout.inspect | Full visual model + report |
layout.overlaps | Elements whose boxes collide |
layout.overflow | Clipped / collapsed / off-screen |
layout.contrast | WCAG AA ratios for text elements |
layout.quality | 0..1 health score + issue list |
layout.fidelity | Which CSS features are exact vs approximated |
3. HTTP API. Call from anything
cornea --serve-http [addr] # default 127.0.0.1:8080
curl -s -X POST http://127.0.0.1:8080/inspect \
-H 'Content-Type: application/json' \
-d '{"html":"<p style=\"color:#cccccc\">bady</p>","width":360}'
{"total_elements":1,"contrast":[{"selector":"...>p","fg":"#cccccc","bg":"#ffffff",
"ratio":1.61,"pass_aa":false}]}
| Route | Method | Returns |
|---|---|---|
/inspect | POST | Full report |
/overlaps | POST | Collisions |
/overflow | POST | Clipped / collapsed / off-screen |
/contrast | POST | WCAG ratios |
/quality | POST | Health score |
/fidelity | GET | Engine capabilities |
/health | GET | Liveness |
Watching a live page (URL capture)
Every surface accepts a URL where the HTML would go. Cornea fetches the page, inlines its external stylesheets and external scripts (relative URLs resolved against the page), then inspects what a browser would actually show. That is the live coding session loop: run your dev server, point cornea at it, read the layout verdict.
cornea http://localhost:3000 390
{ "url": "http://localhost:3000", "width": 390, "height": 844 } // HTTP /inspect
{ "method": "tools/call", "params": { "name": "layout.quality",
"arguments": { "url": "http://localhost:3000", "width": 390 } } } // MCP
Honesty around capture:
- Failed fetches leave the original tag in place and record a note, so the report warnings still flag what did not load.
- A positive
heightemulates a fixed viewport (screenshot frame, iframe, email) and enables below the fold clipping checks. Default 0 means an unbounded scrolling page. - Capture is a snapshot in time. Determinism holds engine side: the same fetched bytes always produce the same report.
- Plain HTTP is supported natively. HTTPS pages need a TLS stack, which the binary does not carry; capture from a local http dev server or pre inline the page with a fetch layer of your own.
Architecture
┌─────────────────────────────────────┐
HTML + CSS ──► │ cornea (one binary) │
│ │
│ dom.rs html5ever ──► tree │
│ css.rs <style> + inline │
│ │
│ layout.rs deterministic layout │
│ (block/inline/flex, │
│ box model, z-index) │
│ │ │
│ model.rs Visual Geometry Model │
│ │ │
│ inspect.rs overlap / overflow / │
│ contrast / quality │
│ │ │
│ rest.rs canonical dispatch │
└───────┬─────────────┬───────────────┘
│ │
┌─────┴────┐ ┌────┴───────────────┐
│ CLI │ │ MCP (stdio) / HTTP│
│ cornea │ │ layout.* / REST │
└──────────┘ └────────────────────┘
Read the full technical spec: CORNEA-ARCHITECTURE.md
Fidelity. Honest about what's approximate
Cornea never silently fakes precision. layout.fidelity tells an agent exactly what it can trust, and every report carries its own warnings for sources the engine saw but did not apply (external stylesheets, external scripts, media queries, unresolved colors):
{
"exact": ["box model", "block flow", "inline text estimates", "flex row/column (no wrap)",
"z-index", "visibility", "absolute/fixed left/top", "inline styles",
"class/id/tag selectors", "WCAG contrast (hex, rgb, hsl, alpha)"],
"approximate": ["text glyph width (not shaping)", "flex-grow/flex-basis distribution",
"percentage widths", "overlap semantics ignore intentional stacking"],
"deferred": ["grid (parsed as block flow)", "media queries", "border-radius",
"external stylesheet <link> when not captured",
"complex selectors (combinators, pseudo)"],
"js": {
"engine": "boa",
"phase": "A",
"enabled": "opt-in via --js / js:true",
"dom_shim": "static HTML mirrored in first; getElementById; innerHTML parses markup",
"unsupported": ["async APIs", "event dispatch", "selector engine", "React/SPA mounting (Phase B)"]
}
}
JavaScript built-in pages (Phase A)
Cornea can execute inline <script> that builds its DOM via a minimal shim,
then run the result through the same deterministic layout engine:
cornea page.html 360 --js # inline scripts build the DOM first
{ "html": "<p>…</p>", "width": 360, "js": true } // HTTP /inspect and MCP layout.*
- Static HTML is mirrored into the shim before scripts run, so scripts can
attach to existing nodes via
document.getElementById, andinnerHTMLparses real markup into elements. Static content survives script runs. style.*assignments are serialized back tostyle="…"and participate in overlap/contrast checks.- Async APIs are rejected (
setTimeout,fetch, …) rather than hung. Any such use is surfaced in the report'sjs_notes, so determinism stays a guarantee. - External
<script src>runs only when a capture layer (URL capture) inlined its body first. - Full React/SPA mounting is Phase B (experimental). Tracked in
ROADMAP-JS.md.
Visual guide: testing (battle-tested)
54 tests run clean with cargo test; CI enforces fmt, clippy -D warnings, release build, tests, and a CLI smoke test on every push.
$ cargo test
Running unittests src/lib.rs ... 36 passed // determinism, overlap, overflow,
Running unittests src/main.rs ... 9 passed // contrast, inline flow, flex,
Running tests/endpoints.rs ... 9 passed // nesting, empty input, MCP, CLI, JS, capture
- Determinism. Same page inspected twice gives byte-identical JSON (the core thesis).
- Bug detection. The fixture's overlaps, overflows, and contrast failures are all asserted.
- End-to-end. The compiled binary is spawned and
layout.*is called over real stdio MCP. - HTTP. The TCP server boots on an ephemeral port and real requests are made.
- Live capture. A real page is served over a local socket; linked CSS must change a contrast verdict.
- Edge cases. Empty HTML, deep nesting (no crash),
display:none, long text, flex row/col, inline wrapping.
Layout support (v1 scope)
| Feature | Status |
|---|---|
| Block flow, box model (content/border-box) | exact |
| Inline text runs (horizontal, wrapping) | exact |
| Flex row / column (simplified, no wrap) | approximate |
| Absolute / fixed positioning (left/top) | exact |
z-index, visibility, display:none | exact |
.class / #id / tag selectors + inline styles | exact |
| WCAG contrast (hex, rgb, hsl, alpha, inherited colors) | exact |
| Live URL capture (CSS and script inlining) | supported |
| Grid, media queries, border-radius | deferred |
Repository layout
cornea/
├── Cargo.toml # crate: html5ever + serde/serde_json, release LTO+strip
├── README.md # this file
├── CORNEA-ARCHITECTURE.md # full technical spec
├── src/
│ ├── lib.rs # build_model / analyze pipeline + unit tests
│ ├── dom.rs # html5ever -> lightweight element tree
│ ├── css.rs # <style> + inline style resolution
│ ├── fetch.rs # live URL capture: GET + css/script inlining
│ ├── layout.rs # deterministic layout engine
│ ├── model.rs # VisualModel / ElementView / Rect
│ ├── inspect.rs # overlap / overflow / contrast / quality / warnings
│ ├── rest.rs # canonical endpoint dispatch (shared by all surfaces)
│ ├── main.rs # CLI + stdio MCP server
│ └── server_http.rs # dependency-free HTTP/1.1 API
├── tests/
│ ├── endpoints.rs # end-to-end CLI + MCP binary tests
│ └── fixtures/sample-bugs.html # known-bug fixture for CI smoke
└── .github/workflows/ci.yml
Status
Working MVP, hardened and battle-tested. Deterministic inspection engine with CLI + MCP + HTTP API, live URL capture, below the fold checks, report warnings, 54 passing tests, green CI, plus Phase A inline-script rendering (--js). The roadmap builds toward giving Cornea (and the sibling Crayon text-to-image project) an even richer perception over subsequent phases.
Publishing
Cornea is distributed four ways: npm (npm install -g optrex, gives the cornea command), crates.io (cargo install cornea), GitHub Release + Homebrew (brew install), and the MCP Registry (layout.* tools discoverable by agents). All are staged in this repo. See PUBLISHING.md for the tokens, server.json manifest, and the release/tag recipe.
mcp-name: io.github.AbduljabbarBXR/cornea
License
MIT. See LICENSE.