ReftrixMCP

Web design knowledge base platform -- layout analysis, motion detection, and quality evaluation via MCP tools.

For frontend engineers, designers, and AI-agent builders who want to analyze real websites and retrieve reusable UI patterns via Claude or any MCP client.

What it does

• Layout analysis -- auto-detect sections (hero, feature, CTA, etc.), extract grid/typography, and generate React/Vue/HTML code
• Motion detection -- discover CSS/JS animations with frame capture (15 px/frame video mode), CLS detection via Pixelmatch
• Quality evaluation -- score designs on three axes (originality, craftsmanship, contextuality) with anti-AI-cliche detection
• Semantic search -- find layout, motion, narrative, background, and responsive patterns via pgvector HNSW hybrid search
• Preference profiling -- learn user design preferences through feedback sessions and personalize search results via reranking (GDPR-compliant)
• Part-level analysis -- extract 16 UI part types (button, icon, heading, etc.) with DINOv2 visual embeddings for visual similarity search
• Vision integration -- Ollama llama3.2-vision for richer layout, motion, and narrative understanding
• Section post-processing -- auto merge/split sections by type, heading, and height (Rule 1-4) for optimal structure
• Multi-tile capture -- split large sections (>viewport height) into tiles for complete DINOv2 visual coverage
• Blank image detection -- detect lazy-loading unrendered sections and re-capture via Playwright for full coverage
• Code generation -- convert analyzed sections to React, Vue, or plain HTML with matched motion patterns
• Unified search -- cross-service search across layout, part, motion, background, and narrative patterns in a single query
• Image similarity search -- find visually similar designs via DINOv2 embeddings from Base64/URL input (RRF 3-source)
• Rate limiting -- Token Bucket + Redis Lua (CWE-770 DoS prevention), 3-tier (analysis 10 RPM / search 120 RPM / default 60 RPM)
• Search cache -- LRU in-memory cache (lru-cache v11) with TTL-based natural expiry (5 min)
• BullMQ UI -- Bull Board dashboard for monitoring async page.analyze jobs (port 21080)
• SBOM -- CycloneDX 1.6 auto-generation for EU CRA vulnerability reporting compliance

Why ReftrixMCP

Quickstart

Run page.analyze on any URL in under 5 minutes.

Prerequisites

Node.js 20+, pnpm 10+, Docker & Docker Compose, Ollama

Setup

git clone https://github.com/TKMD/ReftrixMCP.git && cd ReftrixMCP
pnpm install                                     # CUDA skip is default; see GPU note below
cp .env.example .env.local                       # edit DATABASE_URL / REDIS_URL as needed
cp .env.local packages/database/.env             # Prisma CLI requires this copy
pnpm docker:up                                   # PostgreSQL 18 + pgvector + Redis
pnpm db:migrate && pnpm db:seed
pnpm build
pnpm exec playwright install chromium            # browser for page crawling
pnpm --filter @reftrixmcp/ml download:dinov2        # DINOv2 visual embedding model (~800 MB)
pnpm --filter @reftrixmcp/ml repair:e5-cache --check # (optional) verify multilingual-e5-base ONNX cache (~1.1 GB) integrity
curl -fsSL https://ollama.com/install.sh | sh    # install Ollama
ollama pull llama3.2-vision                      # vision model (~7.9 GB)
ollama serve                                     # keep running in a separate terminal

Note: If you change .env.local, also update packages/database/.env. page.analyze workers are auto-forked by WorkerSupervisor when the MCP server starts (v0.4.0 PR7d-2+). Manual start via pnpm --filter @reftrixmcp/mcp-server worker:start:page is developer-only and requires REFTRIX_ALLOW_MANUAL_WORKER=true to bypass the Redis-based dual-run guard if the MCP server is also running. See Getting Started for GPU configuration and details.

GPU / CUDA: CUDA binary download is skipped by default (CPU fallback). For GPU acceleration setup, see Troubleshooting: CUDA Detection.

Connect to Claude

Add to your MCP config:

• Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
• MCP Client CLI: .mcp.json in the project root or ~/.claude/.mcp.json

{
  "mcpServers": {
    "reftrix": {
      "command": "node",
      "args": ["/absolute/path/to/ReftrixMCP/apps/mcp-server/dist/index.js"],
      "env": {
        "NODE_ENV": "development",
        "DATABASE_URL": "postgresql://reftrix:change_me@localhost:26432/reftrix?schema=public",
        "REDIS_URL": "redis://localhost:27379",
        "OLLAMA_BASE_URL": "http://localhost:11434",
        "OLLAMA_HOST": "http://localhost:11434",
        "ENABLE_SECTION_SCREENSHOT_FALLBACK": "true"
      }
    }
  }
}

Replace change_me with a secure password. Port 26432 = standard 5432 + 21000 offset.

OLLAMA_BASE_URL is used by the MCP server process; OLLAMA_HOST is used by the worker process. Both must match if Ollama runs on a non-default port.

ENABLE_SECTION_SCREENSHOT_FALLBACK enables Playwright-based individual section screenshots for sections outside the initial screenshot range (WebGL/lazy-rendered pages). This significantly improves DINOv2 visual embedding coverage. Set to "false" to disable.

Optional environment variables (defaults work out of the box): MAX_TILES_PER_SECTION (default 20, max 100) -- max tiles per section for multi-tile capture. BLANK_IMAGE_STDDEV_THRESHOLD (default 5.0) -- stddev threshold for blank image detection. DUPLICATE_VECTOR_THRESHOLD (default 0.995) -- cosine similarity threshold for vision embedding dedup. EMBEDDING_IDLE_TIMEOUT_MS (default 30000) -- ONNX Worker VRAM auto-release timer (0 to disable). DINOV2_MODEL_PATH -- custom DINOv2 ViT-B/14 ONNX model path.

Example tools

ReftrixMCP provides 39 MCP tools. Key examples:

• layout.ingest -- fetch a web page, take a screenshot, and extract section patterns
• layout.search -- semantic search over layout sections by natural-language query
• motion.detect -- detect CSS/JS animations with video-mode frame capture
• quality.evaluate -- score design quality on originality, craftsmanship, and contextuality
• page.analyze -- unified analysis: layout + motion + quality + responsive in one call (async via BullMQ), with opt-in Phase 7.5: accessibility audit, performance evaluation, and auto snapshot
• responsive.search -- search responsive analysis results by viewport and breakpoint
• preference.hear -- interactive preference hearing sessions with sample presentation and feedback collection
• preference.get -- retrieve preference profiles (with GDPR data portability support)
• preference.reset -- reset or permanently delete preference profiles (GDPR Right to Erasure)
• part.search -- semantic search over UI parts with visual (DINOv2) or text embeddings
• part.inspect -- get detailed part info including computed styles, bounding box, and accessibility
• part.compare -- compare 2-5 parts side by side on styles, layout, and interaction

Full tool reference: MCP Tools Guide

Architecture

MCP Client (Claude Desktop / Code)  --stdio-->  MCP Server (<!-- gen:tool-count -->39<!-- /gen:tool-count --> tools, Zod)
  +-- Service Layer: Playwright, Sharp+Pixelmatch, DOMPurify
  +-- ML Layer: ONNX Runtime (multilingual-e5-base + DINOv2 ViT-B/14, 768-dim)
  +-- BullMQ Workers: page.analyze, quality.evaluate
  +-- PostgreSQL 18 + pgvector 0.8 (HNSW, tsvector)  +  Redis 7

Documentation

Known limitations

• onnxruntime-node is an optional dependency; ML features (embedding, visual search) require explicit install: pnpm add onnxruntime-node. Non-ML tools (layout analysis, quality evaluation, code generation) work without it
• CPU-mode embedding takes ~2-5 s per text; GPU recommended for batch workloads
• Minimum 16 GB RAM; 32 GB recommended for concurrent analysis with Ollama Vision
• First embedding call downloads ~1.1 GB ONNX model (multilingual-e5-base, FP32) into the transformers.js cache. Verify integrity at any time with pnpm --filter @reftrixmcp/ml repair:e5-cache --check; pass --repair to re-download on size/SHA-256 mismatch, or --force to always re-download
• page.analyze workers are auto-forked by WorkerSupervisor when the MCP server starts (v0.4.0 PR7d-2+); manual start is developer-only (REFTRIX_ALLOW_MANUAL_WORKER=true required when MCP server is running)
• Vision analysis (layout, motion, narrative) requires Ollama + llama3.2-vision running locally
• DINOv2 visual embedding model requires ~800 MB download (ViT-B/14 ONNX)

Release notes / リリースノート

• Plan v4.4 PR-N (2026-05-17): WorkerSupervisorOptions.restartDelayMs field formal removal + env-only canonical SSOT consolidation per ADR-0035 Amendment 1 §Decision 5. The WORKER_RESTART_DELAY_MS and EMBEDDING_BACKFILL_RESTART_DELAY_MS environment variables are now the sole source of truth for per-type restart cooldown values; resolution is performed via getRestartDelayMsForType(workerType). Server version bumped to 0.6.0. / WorkerSupervisorOptions.restartDelayMs フィールドを正式削除し、ADR-0035 Amendment 1 §Decision 5 に従い env-only canonical SSOT へ一元化。WORKER_RESTART_DELAY_MS と EMBEDDING_BACKFILL_RESTART_DELAY_MS 環境変数が per-type restart cooldown 値の唯一の真実源となり、getRestartDelayMsForType(workerType) 経由で解決される。サーバーバージョンを 0.6.0 に bump。

License

AGPL-3.0-only -- see LICENSE.

Network use requires source disclosure per Section 13. Source: github.com/TKMD/ReftrixMCP Commercial license: licence@reftrix.io

Contributing

See CONTRIBUTING.md.

Security

Report vulnerabilities per SECURITY.md. Privacy: docs/legal/PRIVACY_POLICY.md | Profiling privacy: apps/mcp-server/PRIVACY.md | Data retention: apps/mcp-server/DATA_RETENTION.md | Third-party licenses: THIRDPARTY_LICENSES.md