This package is maintained in the AccessLint monorepo.

@accesslint/mcp

MCP server for accessible agentic coding — WCAG audit tools for AI coding agents. Built on @accesslint/core. From AccessLint.

Setup

Add to your MCP client configuration:

{
  "mcpServers": {
    "accesslint": {
      "command": "npx",
      "args": ["@accesslint/mcp"]
    }
  }
}

Related MCP server: A11y MCP Server

Tools

• audit_html — Audit an HTML string for WCAG violations. Auto-detects fragments vs full documents.

• audit_file — Read an HTML file from disk and audit it.

• audit_url — Fetch a URL and audit the returned HTML.

• diff_html — Audit new HTML and diff against a previously named audit to verify fixes.

• list_rules — List available WCAG rules with optional filters by category, level, fixability, or criterion. All audit and diff tools accept an optional min_impact parameter to filter results by severity. Valid values, from most to least severe: critical, serious, moderate, minor. When set, only violations at that level or above are shown.

Each violation in the audit output includes the rule ID, CSS selector, failing HTML, impact level, and — where available — a concrete fix suggestion, fixability rating, and guidance. When multiple elements break the same rule, shared metadata is printed once to keep output compact.

Prompts

React Component Auditing

To audit React components (.jsx/.tsx), the agent uses the audit-react-component prompt, which guides it through:

1. Reading the component source

2. Mentally rendering it to static HTML (acting as renderToStaticMarkup)

3. Passing the rendered HTML to audit_html with component_mode: true

No extra runtime dependencies are required — the agent renders the component itself based on the source code.

Why use this instead of prompting alone?

Without tools, the agent reasons about WCAG rules from memory. The MCP replaces that with structured output — specific rule IDs, CSS selectors, and fix suggestions — so the agent skips straight to applying fixes. This means 23% fewer output tokens per run, which translates directly to faster and cheaper completions.

Benchmarked across 25 test cases, 67 fixable violations, 3 runs each (Claude Opus):

License

MIT