Skip to main content
Glama
StyleTrace

StyleTrace

StyleTrace

English | 繁體中文 | 日本語 | 한국어 | Français | Español

Node >=20 TypeScript Playwright MCP npm downloads

StyleTrace is an MCP server that analyzes references and returns a prompt-ready design brief for agents and reviewers. One use case is helping developers build websites without getting distracted by design decisions, then review generated output against the original style constraints.

StyleTrace reverse engineering comparison

Additional example using the same references with and without StyleTrace in the review flow:

StyleTrace with and without comparison

Why Use It

  • turn a few reference pages, screenshots, or HTML snippets into a prompt-ready design brief an agent can actually use

  • make website regeneration less generic by preserving the parts that feel distinctive

  • review generated HTML or screenshots against the extracted style constraints instead of relying on vague visual judgment

  • analyze only the exact URLs you give it, so the result stays predictable and reviewable

Related MCP server: mcp-mirage-brand-extract

Installation

Requirements:

  • Node.js >=20

  • Playwright Chromium

Install from npm:

npm install -g @agenticbridge/style-trace
npx playwright install chromium

Or run from a local clone:

npm install
npx playwright install chromium
npm run build

Usage

Connect it from your MCP client.

Published package:

{
  "mcpServers": {
    "style-trace": {
      "command": "npx",
      "args": ["-y", "@agenticbridge/style-trace"]
    }
  }
}

Local clone:

{
  "mcpServers": {
    "style-trace": {
      "command": "node",
      "args": ["/absolute/path/to/style-trace/dist/src/index.js"]
    }
  }
}

The server exposes two tools:

  • analyze_website_style

  • review_generated_style

analyze_website_style accepts exact website URLs:

{
  "urls": ["https://www.apple.com", "https://www.framer.com"],
  "targetArtifact": "landing-page",
  "fidelity": "high"
}

It also accepts mixed references:

{
  "references": [
    { "type": "url", "value": "https://www.apple.com/iphone/" },
    { "type": "image", "value": "https://example-cdn.com/reference/hero-shot.png" },
    { "type": "screenshot", "value": "https://example-cdn.com/reference/hero-capture.png" },
    { "type": "html", "value": "<main><section><h1>Hero</h1></section></main>" }
  ],
  "targetArtifact": "prototype",
  "fidelity": "medium",
  "designIntent": "preserve the hero hierarchy and chrome discipline",
  "evidenceMode": "inline"
}

urls remains supported for website-only input. Use references when you want to mix website, image, screenshot, and bounded HTML sources in one request.

The result now includes prompt-ready fields such as visualVocabulary, styleInvariants, styleRisks, softGuesses, compositionBlueprint, variationAxes, blendModes, promptReadyBrief, reviewContract, and originalityBoundary.

review_generated_style checks generated HTML or a generated image URL against a StyleTrace result:

{
  "styleResult": { "...": "StyleTrace analyze_website_style output" },
  "generatedHtml": "<!doctype html><html>...</html>",
  "viewportWidth": 1440,
  "viewportHeight": 900
}

It returns matched invariants, violated constraints, drift notes, and review confidence.

How It Works

analyze_website_style visits exactly the public website URLs you provide with Playwright, and it can also analyze direct public image URLs, screenshot references, and bounded HTML snippets. It extracts narrow, reviewable signals such as module structure, hero treatment, CTA patterns, proof modules, imagery, forms, breakpoints, and signature motifs, then compiles them into a prompt-ready design brief with hard constraints, drift risks, composition structure, and review checks. It does not crawl additional pages, and it does not try to invent a new design system or make speculative recommendations.

review_generated_style runs the generated artifact back through the same lens and compares it to the extracted style contract. The goal is to make style review explicit: what matched, what drifted, and what likely became generic.

Limits

  • public http and https URLs only

  • image and screenshot references must point to direct public image assets such as .png, .jpg, .webp, .gif, .avif, or .svg

  • HTML references are bounded snippets, not full browsing sessions

  • image-only or screenshot-only references produce weaker inference for typography, navigation, forms, motion, and breakpoints than live website references

  • no auth flows or private-network targets

  • stdio transport only

  • no persistence, queueing, or web UI

Contributing and Testing

Run the local checks:

npm run typecheck
npm run build
npm test

For a real MCP transport smoke test:

npm run test:mcp-cli

For the full review artifact flow with source captures, with MCP vs without MCP LLM regeneration, and a composite diff board:

built-in comparison set:

npm run test:e2e -- --instance apple-pixel-samsung
npm run test:e2e -- --instance figma-framer-webflow

Or run it against your own public URLs:

bash scripts/test-mcp-cli.sh https://www.apple.com https://www.framer.com

License

MIT

Available Tools

2 tools
analyze_website_styleAnalyze website styleA
Read-only

Analyze exact public website URLs, public image URLs, or a mix of both and extract a compact design grammar. StyleTrace analyzes only the references you provide. Evidence can be omitted, exported to a sidecar file, or inlined.

Inputs: urls, fidelity, references, designIntent, +2 more. Returns: sites, guideline, synthesis, blendModes, +10 more

review_generated_styleReview generated styleA
Read-only

Review generated HTML or a generated image URL against a StyleTrace result, checking invariant matches, drift, and likely style violations.

Inputs: styleResult (required), generatedHtml, viewportWidth, viewportHeight, +1 more. Returns: mode, viewport, confidence, driftNotes, +3 more

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 2 tools

    Disambiguation5/5

    The two tools have clearly distinct purposes: one extracts design grammar from references, the other verifies generated output against an extracted style. There is no overlap or ambiguity between them.

    Naming Consistency5/5

    Both tool names follow the same verb_noun pattern in snake_case: 'analyze_website_style' and 'review_generated_style'. The naming is consistent and predictable.

    Tool Count3/5

    With only 2 tools, the set is on the thin side. For a server focused on style analysis and review, this minimal set covers the core workflow but feels slightly sparse compared to typical servers.

    Completeness5/5

    The two tools provide a complete lifecycle for the intended domain: analyze a style from references, then review generated output against that style. There are no obvious gaps in the stated workflow.

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      A
      quality
      D
      maintenance
      Provides comprehensive design principles and best practices to help LLMs generate modern, accessible web pages through guidance on layouts, colors, and typography. It enables users to review design approaches and access expert recommendations for responsive design, component structure, and current industry trends.
      12
      37 npm
      3
      -
    • A
      license
      A
      quality
      C
      maintenance
      Analyzes and extracts design tokens, assets, and layout from live websites to enable AI clients to faithfully replicate them, with tools for screenshotting, component inspection, and pixel-diff verification.
      14
      18 npm
      4
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Converts any rendered web page into a complete, LLM-readable design specification—including block screenshots, distilled computed CSS, design tokens, responsive deltas, and alpha-channel-analyzed assets—so models can accurately reproduce the design.
      MIT