Agentic skill · WCAG 2.2 AA · MIT

A sentinel for the accessibility your agent can't see.

A free, open-source skill for AI coding agents. Load it into Claude Code, Cursor, Codex, or Gemini, and your agent checks every screen it builds against all 55 WCAG 2.2 AA criteria, then writes the fix back to source.

Install via your agent 1 line $npx skills add a11y-sentinel

a11y-sentinel · home.tsx3 findings

Image missing alt textCritical · WCAG 1.1.1

Hero.tsx:42

Insufficient contrast (3.1:1)Serious · WCAG 1.4.3

<p class="text-zinc-400 bg-white">…

Pricing.tsx:118

Link text “click here” isn't descriptiveModerate · WCAG 2.4.4

<a href="/docs">click here</a>

Footer.tsx:27

Each finding maps to a one-line fixYour agent, using a11y-sentinel

Claude CodeCursorCodexGemini CLIany Agent Skills tool

Your agent already ships UI. Now it ships accessible UI.

a11y-sentinel rides along in your agent's context. Before it commits a screen, every criterion is already on hand, so accessibility becomes part of the build instead of an afterthought.

Complete. All 55 WCAG 2.2 A/AA criteria, not the handful a model happens to remember.

Judgment-first. It checks what a scanner cannot: whether alt text is meaningful, whether focus order is logical.

Patch-shaped. Each finding comes with a concrete fix, scoped to one file.

Anywhere. Plain Markdown, so it works in any agent that reads skills.

AGENT · SESSION

UPolish the pricing page and ship it.

ABefore I commit, I'll audit Pricing.tsx against WCAG 2.2 AA using a11y-sentinel.

AFound 2 critical and 1 serious. Applying the fixes:

- <img src="/pro.png" />
+ <img src="/pro.png" alt="Pro plan, unlimited seats" />
- <p class="text-zinc-400">
+ <p class="text-zinc-600">

ARe-checked. Contrast and alt text pass. Ready to commit?

From a screen to a patch in your source.

01Audit

Run both passes

The skill walks your agent through the automated pass (axe, pa11y) and the manual one, page by page, against all 55 criteria.

02Judge

Map each finding

Every issue ties to a WCAG criterion and a real fix, explained the way a senior reviewer would. No raw scanner dump in the context window.

03Fix

Write it back

Your agent patches the right file by priority, transaction-blockers first. You review the diff. Nothing changes without your say.

What a linter and a scanner leave out.

axe and eslint catch a real but partial slice. a11y-sentinel hands your agent the whole picture and the fix.

Capability	eslint-a11y	axe-core	a11y-sentinel
All 55 WCAG 2.2 criteria	partial	partial	●
Judgment (meaningful alt, focus order)	no	no	●
A fix for each finding	no	no	●
US ADA / legal context	no	no	●
Works inside any agent	no	no	●

The rules it actually checks.

Each finding maps to a specific WCAG 2.2 criterion and the ADA expectation behind it. No vague “improve accessibility” notes. Every flag points to a rule and a fix.

1.1.1

Non-text content

alt text, decorative imagery, SVG titles

1.3.1

Info & relationships

landmarks, lists, table headers

1.4.3

Contrast (minimum)

text vs. background, focus rings

2.1.1

Keyboard

tab order, escape paths, focus traps

2.4.4

Link purpose

ambiguous link text, button-as-link misuse

2.5.8

Target size

tap targets at least 24 by 24 CSS px

3.3.2

Labels or instructions

form fields, required, error messages

4.1.2

Name, role, value

ARIA misuse, custom widget contracts

Frequently asked.

Does this replace a human accessibility audit?

No. Automated checks cover roughly a third of WCAG. The skill pushes your agent through the manual pass too, but cognitive load, screen-reader UX, and real assistive-tech testing still need a person.

What does the skill actually contain?

Plain Markdown: all 55 WCAG 2.2 A/AA criteria with what to check and how each fails, the US legal context, and a step-by-step audit procedure that leans on axe and pa11y for detection.

Will my agent change code without asking?

That depends on how you run it. The skill tells your agent to fix the source by priority and report honestly. You review the diffs before anything lands.

Is it free?

Yes. MIT, open source, no accounts, no tiers, forever.

How is it different from eslint-jsx-a11y or axe?

Those detect a subset of issues from source or the DOM. The skill gives your agent the full criteria, the judgment layer those tools can't do, a fix for each finding, and the legal context, then walks it through both passes.