Power your agent
The publisher's own command. Not a PlugMCP endpoint.
Host a copy with our keys and logs.
npx -y [email protected]Claude Code has no install link for a local command. Copy the command below and run it in a terminal.
claude mcp add a11yscope-mcp -- npx -y a11yscope-mcp@0.3.0Run in a terminal.
{ "mcpServers": { "a11yscope-mcp": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}claude_desktop_config.json
Remote URL or a command on your machine? How the two installs differ
Run on PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
LaunchFrom the publisher
A Chrome extension that checks the current page against the automatable subset of WCAG 2.2 Level AA, and explains each finding in terms a developer can act on.
Built to stay honest about its own limits: automated tooling catches roughly a third of real accessibility barriers, and the UI says so rather than implying a clean scan means compliance.
Two ways to run the same engine:
MCP server that checks web pages and HTML against WCAG 2.2 AA — the A11yScope engine for AI coding agents
| Install | For | |
|---|---|---|
| Chrome extension | Chrome Web Store | People: scan the page you are looking at, in a side panel (English, 한국어) |
MCP server a11yscope-mcp | claude mcp add a11yscope -- npx -y a11yscope-mcp | AI coding agents (Claude Code, Cursor, Claude Desktop): scan a URL, a file or the HTML the agent just wrote, then fix it. See mcp/. |
Both run locally. Nothing is uploaded, no account, no telemetry.
These are not incidental — they shape every decision in the codebase.
npm is used only for the test runner.activeTab, not host permissions. The extension can read a page only
after the user explicitly scans it. This keeps the install-time permission
warning minimal, which matters for both review and conversion.SHIP comes from node_modules.manifest.json MV3 manifest
src/lib/color.js contrast maths, alpha compositing
src/lib/dom.js visibility, accessible names, selectors
src/rules/ the checks, grouped by theme
src/content/engine.js runs every rule, returns a serialisable report
src/content/highlight.js injected highlighter (serialised into the page)
src/background/ service worker: tab access and injection
src/panel/ side panel UI, report export
tools/make_icons.py draws the PNG icons
tools/package.mjs builds the Web Store zip from an allow-list
test/ fixtures and the Chromium-backed test runner
31 checks across four groups. Each reports either a violation (the failure is unambiguous) or a review item (a real question a human has to answer — never a guess dressed up as a result).
Where a value cannot be determined — text over a background image, a
::placeholder colour, a control hidden off-screen until focused — the rule
reports it for review instead of inventing a number. A scanner that cries wolf
on correct markup is worse than no scanner.
Some rules only ever ask for review, because the standard itself leaves room for
judgement. link-generic-text is the clearest case: WCAG 2.4.4 is Link Purpose
In Context, so a "read more" whose meaning comes from the surrounding card
conforms. Demanding that link text stand alone is 2.4.9, which is Level AAA.
Precision is checked against real sites, not just the fixtures. gov.uk is the
control: it is a well-known example of careful accessibility work, so anything
we report there is a false positive until proven otherwise.
node tools/scan.mjs https://www.gov.uk # expect 0 violations
node tools/scan.mjs https://stripe.com --review
Three real defects were found this way, each of which would have produced failures across most of the web:
text-indent: -5000px + overflow: hidden was measured for
contrast. That is the standard way to give an icon button a screen reader
label. The text is never painted, so its contrast is meaningless. The engine
now measures the text node with a Range and checks it survives every clipping
ancestor.aria-label on descendants. An icon link built
as <a><svg role="img" aria-label="Nvidia"></svg></a> is correctly named, but
a walker that only collects text nodes sees nothing. This alone accounted for
30 bogus "link has no discernible text" findings on stripe.com.Each fix has a regression case in test/clean.html.
<iframe> is not scanned.mix-blend-mode and gradient text. Contrast over a gradient is reported
for review; text composited with a blend mode may still produce a ratio that
does not match what is on screen.npm install # puppeteer, for the tests only
npm test # rule engine against both fixtures
node test/extension.mjs # loads the real extension into Chromium
npm run icons # regenerate PNGs from tools/make_icons.py
npm run package # build dist/a11yscope-<version>.zip
chrome://extensionstest/fixture.html is a deliberately broken page; every rule in MUST_FIRE
has to find something in it. test/clean.html is correct markup and must
produce zero violations — that second assertion is the one that keeps the
tool trustworthy, and it has already caught one false positive during
development.
MCP server that checks web pages and HTML against WCAG 2.2 AA — the A11yScope engine for AI coding agents
{ "mcpServers": { "a11yscope-mcp": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}Tools
No tool list yet. Remote servers get one after the first health check.
FAQ
Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.
No required environment variable is listed on this record.
A11yscope runs locally over stdio.
Last checked 2026-11-05. MCP server that checks web pages and HTML against WCAG 2.2 AA — the A11yScope engine for AI coding agents