PixelJury

PixelJury is a visual QA CLI for AI-built frontends that renders your page, catches generic design and accessibility problems, and writes the fix prompt your coding agent can apply.

Reviews the actual rendered page in Chromium, not just source files or component names.
Flags hard failures developers can fix immediately: contrast, mobile overflow at 390px, small touch targets, tiny text, and repeated AI UI tropes.
Scores typography, hierarchy, color, spacing, originality, and polish with your chosen vision provider.
Writes critique.md, fix-prompt.md, screenshots, and score.json so an agent can run a review -> fix -> re-review loop.

Install

Run it directly:

npx pixeljury review http://localhost:3000

Or install it in a project:

npm install --save-dev pixeljury
pnpm add -D pixeljury
yarn add -D pixeljury
bun add -d pixeljury

Quickstart

Start your app, then point PixelJury at the local URL:

npx pixeljury review http://localhost:3000 --provider mock

That creates:

pixeljury/
  screenshot.png
  screenshot-390.png
  critique.md
  fix-prompt.md
  score.json

Open pixeljury/fix-prompt.md, give it to your coding agent, apply the fixes, and run the same command again. Use a model-backed provider when you want the full visual score:

npx pixeljury review http://localhost:3000 --provider anthropic
npx pixeljury review http://localhost:3000 --provider openai
npx pixeljury review http://localhost:3000 --provider claude-code
npx pixeljury review http://localhost:3000 --provider codex

What You Can Do With It

Review a landing page before shipping it.
Catch the common AI-generated UI tells: gradient heroes, emoji feature cards, fake metrics, repeated card grids, default fonts, and low-contrast footer text.
Give Cursor, Claude Code, Codex, Windsurf, or another coding agent a concrete design-fix prompt instead of vague feedback.
Keep before/after proof in CI or pull requests with screenshots and score.json.
Tune design work against an open rubric instead of a private taste call.

Why PixelJury?

It judges pixels, not intentions. PixelJury opens the URL in Chromium and reviews the page that users will actually see.
Hard fails are deterministic. Contrast, mobile overflow, touch-target size, tiny text, and named slop-trope detections run without an API key.
The output is agent-ready. The generated fix-prompt.md prioritizes hard fails first so a coding agent can make targeted changes.
No hosted service is required. Use provider API keys, local Ollama, Claude Code, Codex, or the deterministic mock provider.
The rubric is public. Scores trace back to rubric.md, so the standard can be inspected and changed.

Compared To Alternatives

Need	PixelJury	Common alternative
Accessibility smoke check	Checks contrast, mobile overflow, touch targets, and tiny text while also producing design feedback.	Axe/Lighthouse are stronger for broad accessibility coverage and performance auditing.
AI UI cleanup	Detects repeated cards, default fonts, gradient panels, fake stats, and generic AI copy in the rendered page.	Design-system rules or prompt snippets help before generation but do not inspect final pixels.
Agent handoff	Writes `fix-prompt.md` with concrete remediation instructions.	Manual design review often leaves feedback in screenshots, comments, or chat.
Visual scoring	Combines deterministic findings with a vision-model rubric.	Pure linters are more deterministic but cannot judge originality, hierarchy, or polish.

Use Lighthouse, Axe, Playwright tests, and visual regression tools when you need their specific coverage. Use PixelJury when the page works but still looks generic, broken, or unready.

Example Output

PixelJury  ·  rubric v0.1

Design score: 61/100   - Generic AI output

Hard fails
  x Page is 70px wider than the 390px viewport (horizontal scroll on mobile)   caps at 70
  x Text contrast 3.1:1 (needs 4.5:1) - e.g. "Copyright 2026 FlowSync..."     caps at 65

Problems
  x "inter" is the primary typeface - an AI/template default        -5
  x Large gradient background panel (1440x420px)                    -5
  x The same card pattern is reused 5x - reads as templated         -5
  x Pill/badge with AI-marketing copy                               -3

-> pixeljury/critique.md   pixeljury/fix-prompt.md   pixeljury/screenshot.png

See It Work

The repository includes a before/after SaaS page pair in examples/. PixelJury rendered and scored both pages.

Before: `examples/sloppy-saas`	After: `examples/polished-saas`

16 / 100 - Broken or pure slop	84 / 100 - Good, fixable gaps
3 hard fails, 6 slop-trope deductions	0 hard fails, 0 deductions

Full scoring artifacts: before.score.json and after.score.json.

Providers

npx pixeljury review <url> --provider anthropic     # ANTHROPIC_API_KEY
npx pixeljury review <url> --provider openai        # OPENAI_API_KEY
npx pixeljury review <url> --provider gemini        # GEMINI_API_KEY
npx pixeljury review <url> --provider ollama        # local model, no API key
npx pixeljury review <url> --provider claude-code   # local Claude Code login
npx pixeljury review <url> --provider codex         # local Codex / ChatGPT login
npx pixeljury review <url> --provider mock          # deterministic checks only

API providers read keys from --key, then PIXELJURY_KEY, then the provider's standard environment variable. Prefer environment variables for keys; command-line arguments can be visible in shell history and process lists. When no provider is selected and no key is available, PixelJury auto-selects mock; if you explicitly choose an API provider, you must provide that provider's key. The deterministic hard-fail and trope checks still run with mock.

--provider claude-code and --provider codex shell out to the local CLI you already have logged in. PixelJury stages screenshots in a temporary working directory for those passthrough runs so the local agent is not pointed at your project checkout. Use API-key providers for shared automation or public CI.

How It Works

Render with Playwright
  -> analyze deterministic static signals
  -> score screenshots with a vision provider
  -> compose the final score and write artifacts

The deterministic stage finds:

WCAG contrast failures
mobile overflow at 390px
touch targets below 44px
tiny body text
overused default fonts
large gradient panels
repeated card patterns
AI-marketing badges and fake-looking stats

The vision stage scores six rubric dimensions: typography, hierarchy, color, spacing, originality, and polish. The final score is capped by hard fails and reduced by deterministic deductions.

API Packages

Most developers should install pixeljury, the CLI. The lower-level packages are published for integrations:

pixeljury-core: Playwright rendering, deterministic static signals, contrast helpers, and score composition.
pixeljury-vision: OpenAI, Anthropic, Gemini, Ollama, Claude Code, Codex, and mock vision-scoring adapters.

Compatibility And Limitations

Requires Node.js 18 or newer.
Uses Playwright and Chromium to render pages. If Chromium is not installed, run npx playwright install chromium.
Reviews a URL, so your app must already be running.
The mock provider runs deterministic checks but does not produce a model-backed visual critique.
PixelJury is not a replacement for full accessibility, performance, security, or visual regression testing.

Development

npm install
npm test
npx playwright install chromium
node packages/cli/bin/pixeljury.js review http://localhost:3000 --provider mock

Workspace layout:

packages/cli      # npx entrypoint and review command
packages/core     # Playwright render, static signals, score composer
packages/vision   # provider adapters and structured-output parsing
examples/         # before/after pages and gallery artifacts
rubric.md         # scoring rubric

pixeljury