VertaaUX MCP Server

The only MCP server with an autonomous audit, fix, and verify loop. Detects UX and accessibility issues across 7 categories, generates framework-aware patches (React, Vue, Angular, Svelte), opens atomic GitHub PRs via the Git Trees API, and verifies the fix landed in production. Built for CI/CD pipelines with policy-as-code thresholds.

Why this server is different

• `verify_fixes` loop: close the audit, fix, re-audit cycle without leaving the agent loop. Budget-capped at 3 iterations to prevent runaway billing.
• Framework-aware patches: suggest_fix detects React/Vue/Angular/Svelte/Nuxt via the nearest package.json and emits idiomatic patches (JSX rewrites for React, HTML attrs preserved elsewhere).
• Atomic Git Trees PRs: generate_pr applies N patches in a single commit or zero. Conflict graph + AST gate (Babel, vue-eslint-parser, svelte/compiler) refuse unparseable patches before they reach the PR.
• Deterministic finding IDs: rule:hash format stable across audit runs so agents can reference findings without storing state.
• Multi-engine a11y: audit_a11y combines axe-core, AccessLint, and VertaaUX analyzers in a single call.
• Policy-as-code: policy_check mirrors the GitHub Action's threshold evaluator exactly so CI and agent verdicts match.

Features

• 38 Tools across audit, fix, PR, schedule, webhook, policy, and a11y categories
• 7 Prompt Templates for common workflows
• 8 Resource URIs for audit data and UX guidelines
• Enterprise Controls: domain allowlist, rate limiting, PII redaction
• Dual Transport: stdio (CLI/Desktop) + HTTP streaming (web)
• Official MCP SDK: spec-compliant via @modelcontextprotocol/sdk

Install

MCP Official Registry

npx -y @modelcontextprotocol/cli install io.github.PetriLahdelma/vertaaux-mcp

npm

npm install -g @vertaaux/mcp-server
VERTAAUX_API_KEY=vx_live_... vertaaux-mcp

Drift policy: smithery.yaml, glama.json, and server.json are auto-generated from the live MCP tool registry by npm run generate:manifests. Never hand-edit them. See docs/REGISTRY-PUBLISHING.md for the runbook.

Quick Start

# Install & build
npm install && npm run build

# Run (stdio transport, for Claude Desktop, VS Code, Cursor)
VERTAAUX_API_KEY=vx_live_... npm start

# Run (HTTP transport, for web clients)
VERTAAUX_API_KEY=vx_live_... npm run start:http

IDE Integration

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "vertaaux": {
      "command": "node",
      "args": ["/path/to/mcp-server/dist/index.js"],
      "env": {
        "VERTAAUX_API_KEY": "vx_live_..."
      }
    }
  }
}

VS Code (with MCP extension)

Add to .vscode/settings.json:

{
  "mcp.servers": {
    "vertaaux": {
      "command": "node",
      "args": ["./mcp-server/dist/index.js"],
      "env": {
        "VERTAAUX_API_KEY": "vx_live_..."
      }
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "vertaaux": {
      "command": "node",
      "args": ["/path/to/mcp-server/dist/index.js"],
      "env": {
        "VERTAAUX_API_KEY": "vx_live_..."
      }
    }
  }
}

Environment Variables

Tools

Audit Tools (Core)

Fix & Verify Tools

Analysis Tools

Management Tools

Accessibility Tools (Multi-Engine)

Deprecated

Prompt Templates

Pre-built workflow prompts for common audit scenarios:

Resources

The server exposes MCP resources via vertaa:// URIs:

Enterprise Controls

Configure domain allowlists, rate limits, and PII redaction programmatically:

import { configureEnterpriseControls } from './server.js';

configureEnterpriseControls({
  allowlist: {
    allowed_domains: ['*.example.com'],
    denied_domains: ['internal.example.com'],
  },
  budget: {
    max_requests: 100,
    max_pages: 50,
    max_duration_ms: 60000,
    max_concurrency: 3,
  },
  redaction: {
    redact_emails: true,
    redact_phone_numbers: true,
    redact_credit_cards: true,
    custom_patterns: [
      { name: 'api_key', pattern: 'sk_[a-zA-Z0-9]{20,}', replacement: '[REDACTED]' }
    ],
  },
});

Example: Audit-to-PR Workflow

1. audit_url({ url: "https://example.com", mode: "deep" })
   → Returns audit_id with top 5 issues

2. get_findings({ audit_id: "...", severity: "critical" })
   → Returns all critical findings with deterministic IDs

3. suggest_fix({ audit_id: "...", finding_id: "button-name:a1b2c3d4" })
   → Returns search/replace patch with 85% confidence

4. run_verification_suite({ url: "...", selector: "button.submit", rule_id: "button-name" })
   → Verifies fix resolves the issue

5. create_pr_comment({ file_path: "src/Button.tsx", patches: [...] })
   → Generates PR comment with suggestion blocks

Development

Project Structure

mcp-server/
├── src/
│   ├── index.ts              # Main entry, tool registration
│   ├── server.ts             # MCP server config, resources, middleware
│   ├── a11y-tools.ts         # Multi-engine a11y audit & baseline diffing tools
│   ├── http.ts               # HTTP transport entry point
│   ├── prompts.ts            # MCP prompt templates
│   ├── analysis.ts           # Component analysis engine
│   ├── patch.ts              # Patch generation
│   ├── verification.ts       # Patch verification
│   ├── pr-comment.ts         # PR comment generation
│   ├── tools/
│   │   ├── audit-url.ts      # audit_url tool
│   │   ├── audit-repo.ts     # audit_repo tool (static analysis)
│   │   ├── audit-artifact.ts # audit_artifact tool (HAR/Lighthouse)
│   │   ├── get-findings.ts   # get_findings tool
│   │   ├── explain-finding.ts# explain_finding tool
│   │   ├── suggest-fix.ts    # suggest_fix tool
│   │   ├── generate-pr.ts    # generate_pr tool
│   │   └── index.ts          # Tool exports
│   ├── transports/
│   │   ├── stdio.ts          # Stdio transport (default)
│   │   └── http.ts           # HTTP streaming transport
│   ├── middleware/
│   │   ├── allowlist.ts      # Domain/path allowlist
│   │   ├── budget.ts         # Rate limiting & quotas
│   │   ├── redaction.ts      # PII redaction
│   │   └── index.ts          # Middleware stack
│   ├── resources/
│   │   ├── audit-results.ts  # vertaa://audits/* resources
│   │   ├── screenshots.ts    # vertaa://screenshots/* resources
│   │   ├── historical.ts     # vertaa://history/* resources
│   │   ├── legacy.ts         # Guidelines resources
│   │   └── index.ts          # Resource exports
│   ├── schemas/
│   │   ├── audit.ts          # Audit schemas (mode, findings)
│   │   ├── findings.ts       # Finding schemas
│   │   ├── controls.ts       # Enterprise control schemas
│   │   ├── errors.ts         # Error schemas
│   │   └── index.ts
│   ├── utils/
│   │   ├── error-recovery.ts # Structured errors with recovery guidance
│   │   ├── change-tracker.ts # Baseline change tracking
│   │   └── deterministic-id.ts # Stable finding IDs
│   └── index.test.ts         # Test suite
├── README.md
├── package.json
├── tsconfig.json
└── vitest.config.ts

Testing

npm test              # Run test suite
npm run test:watch    # Watch mode
npm run test:coverage # Coverage report

Building

npm run build         # TypeScript → dist/

Error Handling

All errors include structured recovery guidance:

{
  "code": "AUDIT_NOT_FOUND",
  "message": "Audit abc123 not found.",
  "recovery": {
    "action": "Start a new audit for this URL",
    "tool": "audit_url",
    "params": { "url": "https://example.com" }
  }
}

Error codes follow JSON-RPC 2.0: -32700 (parse), -32600 (invalid request), -32601 (method not found), -32602 (invalid params), -32603 (internal error).

API Reference

The MCP server communicates with the VertaaUX API v1. See the API Documentation.

License

MIT