Design Atlas
Type
agent skill collection
Verdict
Useful
Licence
Ship licence, open source permissiveopen-source-permissivefree. The README and plugin manifest say MIT, but the repo has no licence file (about 100 stars at review, plugin version 0.10.3, last change 2026-08-25). The npm packages the skills run (@accesslint/cli, @accesslint/mcp, @accesslint/core, @accesslint/chrome) are published as MIT. A separate hosted connector needs an accesslint.com account.
Agent channels
mcp, cli, skill
Pricing
free
Reviewed

← Atlas · Topics: agent-skills, ux-patterns

AccessLint Skills ​

Type
agent skill collection
Verdict
Useful
Licence
Ship licence, open source permissiveopen-source-permissivefree. The README and plugin manifest say MIT, but the repo has no licence file (about 100 stars at review, plugin version 0.10.3, last change 2026-08-25). The npm packages the skills run (@accesslint/cli, @accesslint/mcp, @accesslint/core, @accesslint/chrome) are published as MIT. A separate hosted connector needs an accesslint.com account.
Agent channels
mcp, cli, skill
Pricing
free
Reviewed

What it is ​

AccessLint is a set of five accessibility skills built on the @accesslint/core rule engine. Each skill has one job. accessibility-scan runs the engine on one live page and returns a worklist, locating problems without editing. accessibility-inspect drives the page through what an engine cannot decide: focus order, names and roles, reflow and zoom, reduced motion, form errors and target size. accessibility-audit follows the W3C WCAG-EM method (define scope, explore, sample pages and states, evaluate, report) and gives each criterion a pass, fail or undetermined result. accessibility-fix takes a baseline, edits, and verifies. accessibility-diff reports only the violations a change introduced or fixed. A shared methodology.md sets the honesty rules. The repo also publishes a benchmark in which the same model, with and without the skills, audited trap-laden test pages. Each skill had about 600 to 750 installs on skills.sh.

When to open it ​

  • When you need a real WCAG 2.2 AA audit of a site, with sampling and per-criterion results, not a one-page lint.
  • As a regression check on a branch or on uncommitted changes, locally or in CI.
  • When you want an agent that says "undetermined, needs a screen-reader user" instead of passing a page it could not fully test.

Most useful ​

  • Two grades per finding: severity, and evidence basis (● verified and reproducible, ◐ has evidence but a person must decide, ○ needs assistive technology or lived experience). When unsure between two grades, use the lower one. Findings are marked with symbols, not colour, so the report stays readable for colour-blind readers.
  • Source-grounded worklists: each violation gives the selector, file:line (symbol) when source maps allow, the evidence (such as the measured contrast ratio), and either a mechanical fix or NEEDS HUMAN.
  • Fix without inventing: mechanical fixes are applied as given. Anything that needs alt text, labels or link wording is left as a marked note in the code with the rule ID, and the result is checked by re-running the baseline.
  • Coverage honesty: the methodology cites about 57% of defects found by automation (Deque) and warns not to confuse that with the share of criteria that can be auto-checked.
  • Benchmark result (August 2026): with the skills, the model raised fewer false positives (5 against 11) and in 2 of 3 runs declined to certify a "clean" page it could not fully test (without them, 0 of 3), at about 53% more tokens on a full assessment.

Using it with agents ​

Run npx skills add AccessLint/skills, or in Claude Code claude plugin marketplace add accesslint/skills and then claude plugin install accesslint@accesslint. The plugin also adds the local MCP (audit_live, audit_html, list_rules, explain_rule). Plain requests trigger it ("is localhost:3000 accessible", "fix the a11y issues in Nav.tsx"), as do slash commands such as /accessibility-audit --level AA <url> and /accessibility-diff --branch main. Running npx @accesslint/cli init saves named targets in accesslint.config.json. inspect needs a browser MCP (Chrome DevTools, Playwright or Puppeteer); without one it runs only the static checks.

Watch out for ​

  • The skills run npx -y @accesslint/...@latest on every call, so each run downloads and executes the newest package from npm. Pin versions if that matters to you.
  • scan, audit, fix and diff launch a local Chrome over CDP when none is running, and can reach any URL that Chrome can, including staging and production.
  • accessibility-diff stashes your uncommitted changes to capture a baseline, then restores them. Commit or back up risky work first.
  • The hosted connector (mcp.accesslint.com) is a different service. It stores journeys, runs and findings under an accesslint.com account, and it cannot reach localhost.
  • Licence gap: "MIT" is stated in the README and manifest, but there is no LICENSE file.
  • Small project (about 100 stars) and pre-1.0. Expect changes.

Reusable ideas ​

  • Give every finding a separate evidence grade, and when unsure, use the lower one.
  • Keep locate, assess, fix and guard as separate skills that hand work to each other.
  • Say what kind of check you ran: a single URL without sampling is a scan, not an audit.
  • Write down what must go to a human instead of letting the agent pretend to be a screen-reader user.
  • Test a skill against the same model without it, using fixtures with planted traps, before you ship it.

Web Quality Skills, Anthropic Design Plugin, Inclusive Components, Vercel Web Design Guidelines, Impeccable