A11y Toolkit MCP Server

a11y-toolkit — the accessibility layer for AI coding agents

WCAG 2.2 for your coding agent — audit, fix, document and watch. Zero dependencies, es/en.

32 MCP tools, 6 prompts and a skill covering the whole WCAG 2.2 loop: audit, fix, document, watch, hand off — and now blindfold usability testing, where the agent attempts real tasks perceiving only what a screen reader exposes. 51 of 55 A/AA criteria carry automated signals (93%), every finding ships with a concrete remediation your agent can apply, and the rendered tools run on whichever browser you have — Chromium, Firefox, WebKit, Chrome or Edge.

The European Accessibility Act is in force since June 2025, ADA suits keep landing, and AI agents now write most of the web. A scan is not a defence: the machine finds, the human signs.

Quickstart

# MCP — one line, any MCP client (Claude, Cursor, Codex, Windsurf, …)
claude mcp add a11y-toolkit -- uvx --from a11y-toolkit a11y-toolkit-mcp

# CLI — same engine, one command
pip install a11y-toolkit
a11ytoolkit audit --url https://example.com
a11ytoolkit pair "#1f2328" "#fbfaf7"

Static audits, contrast math, PDF/DOCX screening and the evidence pack run with zero dependencies. The rendered tools (real contrast over images, keyboard traps, reflow, forms) additionally need Playwright: pip install playwright && playwright install chromium.

Features

The tools (32)

Full reference — descriptions, arguments, return shapes and behavior — in docs/tools.md. All tools accept lang (en|es) unless noted; output is JSON unless noted. Generated from server.TOOLS by scripts/gen_readme.py.

The 6 prompts

• audit-page (url*, language): Full accessibility audit workflow for a URL: automated tools + what to verify manually (keyboard, screen reader, zoom). Every prompt exists in English and Spanish.

• fix-contrast (fg*, bg*, context): Guided contrast fix: verify the pair (or the text-over-image region), get the nearest passing color, apply it to the code. Every prompt exists in English and Spanish.

• pre-deploy-check (url*, baseline): Pre-deploy regression check: express audit of the URL plus snapshot diff against a baseline snapshot. Every prompt exists in English and Spanish.

• declaration-eaa (entity*, url*, status): Generate a European Accessibility Act / RD 1112/2018 accessibility statement, collecting any missing legal fields from the user. Every prompt exists in English and Spanish.

• conformance-wcagem (url*, tier, language): Guided WCAG-EM conformance ladder for a site: express screening → agent-verified guided evaluation on a representative sample → conformance report inputs. Every prompt exists in English and Spanish.

• blindfold-task (url*, goal*, kind, language): Blindfold usability test — the agent attempts a real task perceiving ONLY the accessibility tree and acting only by accessible name and keyboard, logging friction per step; closes with a11y_journey_verdict. The capability no other accessibility tool has: task-based non-visual testing where the agent IS the screen-reader user. Every prompt exists in English and Spanish.

Install

mcp-name: io.github.kinti/a11y-toolkit · PyPI: a11y-toolkit · Zenodo DOI: 10.5281/zenodo.22843722

Works with any MCP-capable client — Claude Code/Desktop, Cursor, Windsurf, VS Code, Codex CLI, OpenCode, ZCode, Zed, Cline, Continue, Kimi Code… See docs/clients.md for every verified config format.

claude mcp add a11y-toolkit -- uvx --from a11y-toolkit a11y-toolkit-mcp

Or with JSON config:

{
  "mcpServers": {
    "a11y-toolkit": {
      "command": "uvx",
      "args": ["--from", "a11y-toolkit", "a11y-toolkit-mcp"],
      "timeoutMs": 60000
    }
  }
}

Rendered tools use Playwright if present and accept a browser parameter (auto, chromium, firefox, webkit, chrome, msedge — auto-detects the first available). Accept auth_state (Playwright storage_state path) to audit behind login. Everything else works with zero dependencies.

The skill

git clone https://github.com/kinti/a11y-toolkit && cd a11y-toolkit
./skill/install-skill.sh     # → ~/.zcode/skills and ~/.claude/skills

Agent hook — a11ytoolkit init

Agents that read AGENTS.md on their own get the rule that matters — after any UI change, re-audit before calling the work done — with the exact commands:

cd your-project
uvx --from a11y-toolkit a11ytoolkit init     # writes/merges the AGENTS.md hook (idempotent)

It also prints the MCP config block and the optional skill install step. --check reports whether the hook is present (exit 1 if not — CI-friendly).

Languages — community locale packs

All user-facing messages (findings, remediations, legal wording) live in locale packs — one JSON per language, es and en ship today. Adding a language is a single-file PR: copy _template.json, translate keeping the {placeholders}, and python3 tests/test_locales.py validates it mechanically (keys, placeholders, plural rule, completeness). Same community model as axe-core's locales — with validation axe-core doesn't have. Guide: docs/i18n.md. Downstream products can layer brand wording via overlays without forking.

CLI — same engine, one command

Install it with pip install a11y-toolkit (or uv tool install a11y-toolkit).

a11ytoolkit audit --changed                     # ONLY what changed since HEAD (CI gate)
a11ytoolkit audit --changed --base origin/main  # a branch's changes — exit 2 on high/medium
a11ytoolkit audit --url https://example.com --pages 5   # sitemap-first crawl
a11ytoolkit pair "#1f2328" "#fbfaf7"                    # contrast
a11ytoolkit image hero.jpg --text "#fff" --region 120,40,420,90
a11ytoolkit forms https://mysite/contact                # form errors
a11ytoolkit kbd https://mysite                          # keyboard traps + 3.2.1/3.2.2
a11ytoolkit reflow https://mysite                       # 320px reflow
a11ytoolkit scroll https://medium.com/feed              # infinite scroll
a11ytoolkit validate --url https://example.com          # W3C Nu
a11ytoolkit hover https://mysite                        # tooltip dismissibility
a11ytoolkit fix --file page.html -o fixed.html          # safe autofix
a11ytoolkit declaration --entity "Acme" --url https://… --status partial --framework eaa
a11ytoolkit snapshot https://mysite --out before.json   # before deploy
a11ytoolkit diff before.json after.json                 # after deploy
a11ytoolkit evidence audit.json -o pack.json            # verifiable evidence pack
a11ytoolkit disprove --url https://mysite               # re-verify findings
a11ytoolkit budget --budget budget.json --audit audit.json  # only NEW findings block
a11ytoolkit sarif --from-audit audit.json -o a11y.sarif # GitHub code scanning
a11ytoolkit badge --score 92 --out badge.svg            # honest SVG
a11ytoolkit contrast-dom https://mysite --selector ".hero h1"  # auto-region pixel sampling
a11ytoolkit vpat pack.json --product "Acme" -o vpat.html      # VPAT 2.5 (WCAG) draft
a11ytoolkit earl --from-audit audit.json -o earl.json          # ACT Rules / EARL export
a11ytoolkit wcag3 --from-audit audit.json                      # WCAG 3 draft preview (no score)

Coverage: 51 of 55 WCAG 2.2 A/AA criteria (93%)

The 4 manual-only criteria each have a knowledge entry (a11y_criterion) telling the agent exactly how to verify them by hand. The boundary is printed on every report.

The engine behind a11y.es

a11y.es is a Spanish-market accessibility service that runs this engine. The service is a separate codebase; this repository is the engine.

Validated against real pages

Benchmarked against axe-core 4.10 on real pages monthly (methodology and results). Task-based agent benchmark on live sites, including a stratified sample of Spain's most important pages (method and results). Found a real WCAG failure on gov.uk that axe does not report (blue button at 3.91:1, manually verified). Drove out our own false positives (hidden skip links, honeypot fields, single-context generic links, image-alt accname — each with a regression fixture).

Also run against the official W3C ACT test cases: scripts/act_report.py regenerates the EARL implementation report — 100% consistent with every published ACT rule it implements (355 official test cases, 20/20 rules, zero inconsistencies, verified per release against the w3c/wcag-act-rules repository; per-rule breakdown in docs/act-report-summary.md). The WCAG 2.2 catalog (86 criteria, A/AA/AAA) and the WCAG 3 outcome registry are likewise verified against their canonical w3.org sources (scripts/wcag22_ingest.py, scripts/wcag3_ingest.py).

Why I built this

I have been auditing websites for accessibility since 2003. In twenty years the tools got faster and the failures stayed the same: the same missing alt text, the same keyboard traps, the same forms that swallow your work. When AI agents started writing most of the web, the audit gap stopped being a staffing problem and became an infrastructure one.

So I built the tool I always wanted next to me on an audit. Not another scanner — scanners exist. A toolkit that does half the work, says exactly where it stops, and packages the rest for the person who signs. The gov.uk failure you read about above is not a synthetic benchmark case: this engine found it, and I confirmed it with my own eyes before writing it down.

If it saves you an afternoon, leave a star — it genuinely helps the next auditor find it. If it doesn't, open an issue and tell me what is missing.

Honesty, built in

Every audit says: automation covers 93% of the 55 A/AA criteria (roughly a third of all WCAG once AAA and the genuinely-human checks are counted) — the rest needs a human. The audit-page prompt and the skill have the agent check what it can (keyboard, focus, zoom, announced errors) using the manual checklist. A filter, not a verdict. The remaining human work is priced per criterion — MIN 23 (read and confirm), MED 19 (verify in the browser), MAX 13 (real interaction and judgment); full table. And the evidence pack is the handoff object for the person who signs, screen reader transcript included, so the reviewer sees what a blind user hears.

Security & scope

A local tool: runs on your machine as your user. a11y_audit_url accepts http/https only (no file:// — use --file/html for local HTML). a11y_html_validate in html mode POSTs content to the W3C service (url mode shares only the URL; self-hosted vnu supported). path/output_path read/write local paths — use it in MCP clients you trust.

Development

python3 tests/test_contrast.py && python3 tests/test_audit.py && python3 tests/test_v32.py \
  && python3 tests/test_dom.py && python3 tests/test_cli.py && python3 tests/test_solido.py \
  && python3 tests/test_mcp.py && python3 tests/test_png.py && python3 tests/test_vpat.py \
  && python3 tests/test_act.py

tests/test_dom.py self-skips without Playwright. tests/test_solido.py enforces the design invariants (catalog parity, version alignment, count coverage, hostile-HTML fuzz). Releases: make release V=X.Y.Z — bumps, gates, tags and pushes atomically.

Roadmap

Done recently: rendered audit with shadow DOM and state contrast, form error testing plus autofix, keyboard traps, infinite scroll, reflow, the W3C Nu integration, SARIF and error budgets, the evidence pack (spec published, self-verifying), sitemap crawling, behind-login auditing, browser auto-detection, the screen reader transcript, and per-criterion effort pricing. The full history lives in the changelog.

Next, roughly in this order:

• Attestation interlock: let an external auditor's signature reference the evidence pack by hash (their schema, my side adapts)
• Publish the monthly axe-core benchmark results from CI, run over real pages
• Screen-reader transcript cross-checks against NVDA and VoiceOver output
• More verified client configs (JetBrains AI, Gemini CLI), same treatment as docs/clients.md

Author

Jesús Quintana Fernández (jquin.net) — SEO/GEO consultant and web-accessibility practitioner since 2003. MIT © 2026.