About Vertaaux/mcp Server
VertaaUX MCP Server lets developers and coding agents audit UX and accessibility, generate framework-aware fixes, and verify them. It also supports GitHub PRs and CI policy checks.
Key features
- 01UX and accessibility audits
- 02Framework-aware patch generation
- 03Patch verification loop
- 04Atomic GitHub PR creation
- 05Policy threshold checks
Use cases
- 01Audit a deployed site or local codebase
- 02Generate and verify accessibility fixes
- 03Open a draft PR with patches
- 04Align agent and CI policy results
MCP support
Exposes audit, fix, verification, PR, scheduling, webhook, and policy tools, along with prompts and audit resources, to MCP clients. Supports stdio and HTTP streaming transports.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y @vertaaux/mcp-serverVERTAAUX_API_KEYSecretVertaaUX API key (get one at https://vertaaux.ai/settings/api)Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMEVertaaUX 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_fixesloop: 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_fixdetects React/Vue/Angular/Svelte/Nuxt via the nearestpackage.jsonand emits idiomatic patches (JSX rewrites for React, HTML attrs preserved elsewhere). - Atomic Git Trees PRs:
generate_prapplies 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:hashformat stable across audit runs so agents can reference findings without storing state. - Multi-engine a11y:
audit_a11ycombines axe-core, AccessLint, and VertaaUX analyzers in a single call. - Policy-as-code:
policy_checkmirrors 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, andserver.jsonare auto-generated from the live MCP tool registry bynpm run generate:manifests. Never hand-edit them. Seedocs/REGISTRY-PUBLISHING.mdfor 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
| Variable | Required | Default | Purpose |
|---|---|---|---|
VERTAAUX_API_KEY | Yes | — | API authentication key |
VERTAAUX_API_BASE | No | https://vertaaux.ai/api/v1 | API endpoint URL |
PORT | No | 8787 | HTTP transport port |
GITHUB_TOKEN | No | — | GitHub API access for generate_pr |
Tools
Audit Tools (Core)
| Tool | Description |
|---|---|
audit_url | Run UX & accessibility audit on a deployed URL. Returns top 5 issues with severity breakdown. |
audit_repo | Static analysis on local codebase (React/Vue/Svelte/HTML). Finds missing alt text, unlabeled buttons/inputs/links. |
audit_artifact | Audit from HAR files (response times, failed requests, large payloads) or Lighthouse JSON (accessibility findings). |
get_findings | Retrieve findings from a completed audit with filtering by severity, rule, and pagination. |
get_audit | Get audit job status and results by job ID. |
Fix & Verify Tools
| Tool | Description |
|---|---|
explain_finding | Deep-dive into a finding: WCAG criteria, repro steps, fix guidance, before/after code examples. |
suggest_fix | Generate search/replace patch with confidence score. Supports single and batch mode. |
generate_patch | Generate accessibility fix patch for a specific issue from an audit. |
run_verification_suite | Verify a patch fixes the issue without regressions via before/after audit. |
generate_pr | Create a draft GitHub PR with fix patches. Requires GITHUB_TOKEN. |
create_pr_comment | Generate a PR comment with suggestion blocks, ordered by severity. |
Analysis Tools
| Tool | Description |
|---|---|
analyze_component | Heuristic UX review of component code (no browser needed). Checks images, buttons, inputs, links. |
run_llm_audit | Provider-agnostic LLM audit (Mistral/OpenAI via Vertaa adapter). |
capture_screenshot | Capture screenshot by running a quick audit. |
compare_competitors | Compare UX metrics against competitor URLs with category-level score deltas. |
explain_issue | Format an issue into developer-friendly markdown guidance. |
Management Tools
| Tool | Description |
|---|---|
create_webhook | Register webhook for audit notifications. |
list_webhooks / delete_webhook | Manage webhooks. |
create_schedule | Cron-based scheduled audits with score threshold alerts. |
get_schedule / list_schedules / update_schedule / delete_schedule | Manage schedules. |
get_quota | Check plan and remaining credits. |
get_engines | List available engine versions. |
Accessibility Tools (Multi-Engine)
| Tool | Description |
|---|---|
audit_a11y | Multi-engine accessibility audit using axe-core, AccessLint, and custom analyzers. Returns WCAG-mapped findings with structured fix suggestions and fixability ratings. Supports min_impact filtering and mode (basic/standard/deep). |
diff_a11y | Compare current accessibility findings against a saved baseline. Returns fixed, new, and unchanged findings with net change summary. Requires a prior audit_a11y call to establish the baseline. |
Deprecated
| Tool | Description |
|---|---|
run_audit | DEPRECATED — Use audit_url instead. |
Prompt Templates
Pre-built workflow prompts for common audit scenarios:
| Prompt | Description | Arguments |
|---|---|---|
quick_audit | Audit a URL and summarize top issues with fix recommendations | url |
fix_accessibility | Full audit → patch → PR comment workflow | url |
compare_ux | Compare against competitors and identify UX gaps | url, competitors, industry? |
monitor_regression | Set up scheduled monitoring with alerts | url, frequency? |
audit_codebase | Static analysis on local codebase | path |
Resources
The server exposes MCP resources via vertaa:// URIs:
| URI Pattern | Description |
|---|---|
vertaa://audits/{auditId} | Full audit result |
vertaa://audits/{auditId}/summary | Lightweight summary |
vertaa://audits/{auditId}/findings/{findingId} | Single finding detail |
vertaa://screenshots/{auditId} | Screenshot metadata |
vertaa://screenshots/{auditId}/annotated | Annotated screenshot |
vertaa://history/{encodedUrl} | Audit history for URL |
vertaa://history/{encodedUrl}/trend | Score trend analysis |
vertaa://guidelines/{topic} | UX guidelines (buttons, forms, navigation, color-contrast, errors, content) |
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
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.