@accesslint/mcp

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

This server is a thin adapter: it holds no Chrome, CDP, or audit engine of its own. Live-page audits shell out to the @accesslint/chrome and @accesslint/cli binaries (the same path the AccessLint skills use); HTML-string audits run the @accesslint/cli engine in-process. The server exposes those as MCP tools and formats the results for an agent.

Setup

Add to your MCP client configuration:

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

For live-page audits, audit_live ensures a debuggable Chrome via @accesslint/chrome: it attaches to one already serving the CDP discovery API, or launches a fresh headless one in the background — no manual --remote-debugging-port flag needed. A system Chrome/Chromium (found via CHROME_PATH or a platform search) is used when present; set ACCESSLINT_CHROME_DOWNLOAD=1 to fall back to a managed Chrome for Testing build when none is installed.

Tools

• audit_live — Audit a live URL via CDP. Ensures a debuggable Chrome (auto-launching one headless if none is reachable — no manual setup needed), then runs @accesslint/core against the live DOM. CSP-bypassing. Optional selector scopes the audit to one element; wait_for gates the audit until a selector or text appears. Override the endpoint with port / host (or ACCESSLINT_CDP_PORT) to attach somewhere specific.
• audit_html — Audit an HTML string for WCAG violations (in-process JSDOM). Auto-detects fragments vs full documents; pass component_mode to force fragment handling. Used by the audit-react-component prompt to audit JSX after the agent renders it to a string.
• list_rules — List available WCAG rules with optional filters by category, level, fixability, or criterion.
• explain_rule — Detailed metadata for a single rule (description, WCAG criteria, fixability, browser hint, guidance).

File-on-disk audits go through Read + audit_html. For static-site CI workflows, diffing against a baseline, and snapshot-based regression gates, use the @accesslint/cli package directly (accesslint scan <file-or-url>, --snapshot) or the AccessLint scan / diff skills.

All audit tools accept an optional min_impact parameter to filter by severity. Valid values, from most to least severe: critical, serious, moderate, minor — only violations at that level or above are shown. They also accept rules / wcag allow-lists (run only the listed rule IDs / WCAG criteria) and format: "compact" for one-line-per-violation output.

Each violation in the output includes the rule ID, CSS selector, failing HTML, impact level, and — where available — a fixability rating, browser hint, and remediation guidance. When multiple elements break the same rule, shared metadata is printed once to keep output compact. Live-DOM audits against React dev builds attach a Source: <file>:<line> (Symbol) pointer per violation via React DevTools fibers.

Prompts

React Component Auditing

To audit React components (.jsx/.tsx) without a running app, 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

No extra runtime dependencies are required — the agent renders the component itself based on the source code, and audit_html auto-detects the fragment.

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 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):

| | With @accesslint/mcp | Agent alone | | -------------------- | -------------------- | --------------- | | Violations fixed | 99.5% (200/201) | 93.5% (188/201) | | Regressions | 1.7 / run | 2.0 / run | | Cost | $0.56 / run | $0.62 / run | | Duration | 270s / run | 377s / run | | Timeouts | 0 / 63 tasks | 2 / 63 tasks |

Measured against the v0.8.x in-process architecture. The structured-output advantage (fix rate, tokens) carries over to the current subprocess design; absolute timings will differ.

License

MIT