wcagc-mcp

An MCP server that lets an AI assistant (Claude, ChatGPT, or any MCP-compatible client) run real, deterministic accessibility scans through wcagc — axe-core under the hood, not an LLM guess. Every scan result carries an explicit coverage disclaimer and never claims "compliant": automated testing finds only a portion of accessibility barriers, and this tool says so in every response.

This package is a thin, stateless adapter. It holds no database, no scan logic, and no secrets beyond the wcagc API base URL — it translates MCP tool calls into HTTP calls against the wcagc API and forwards the caller's own bearer. All authentication, entitlements, quotas, and scan orchestration live in the API; this code is safe to read end to end.

Two ways to run it

Local (stdio) — for Claude Desktop, Cursor, or any MCP client that spawns a local process:

npx @wcagc/mcp

Configure your MCP client with:

{
  "mcpServers": {
    "wcagc": {
      "command": "npx",
      "args": ["-y", "@wcagc/mcp"],
      "env": {
        "WCAGC_MCP_KEY": "<your mcp:scan API key>"
      }
    }
  }
}

Mint an mcp:scan key from your wcagc account under Settings → API keys — available on every plan, with a daily quota on Free/Starter and unlimited on Pro/Agency.

Hosted (Streamable HTTP + managed OAuth) — what Claude web/desktop/mobile connectors and ChatGPT use, since neither runs a local process for you. Add this remote MCP connector:

https://mcp.wcagc.com/mcp

The client discovers /.well-known/oauth-protected-resource/mcp, opens the wcagc login/consent flow, and binds the connection to one Organization. No key copy/paste is required. API-key bearer authentication remains supported for local stdio and CI.

ChatGPT availability depends on the ChatGPT plan and on whether the client permits action tools; scan_url creates a scan and is not a read-only operation. See wcagc.com/integrations/mcp.

Tools

One id, one poll tool: whatever scan_url did with a URL, get_scan and get_findings read it back. get_run and get_run_findings are only for full-site runs from scan_site.

Every scan-producing tool returns the coverage disclaimer in both the text content and the structured content. There is no score, grade, or conformance verdict — automated testing finds roughly 30–57% of accessibility issues, and the remainder needs manual review.

Configuration

Development

npm install
npm run dev        # hosted, watch mode
npm run start:stdio # stdio mode
npm run typecheck
npm run verify      # node:test against a local fixture API

License

MIT — see LICENSE.