What is Accessibility Agents?
Accessibility Agents, from the Community Access organisation, is two things in one MIT-licensed package. The first is 113 Agent Skills (six routers that dispatch 107 specialists) that tell Claude Code, Codex, GitHub Copilot, Gemini CLI and ChatGPT how to review web UI, documents and Markdown against WCAG 2.2 AA, plus a hook that blocks edits to user-facing files until the accessibility review has run. The second is an MCP server whose README counts 39 tools: axe-core and Playwright browser scans, scanners for Word, Excel, PowerPoint, PDF, ePub and Markdown, veraPDF validation, contrast maths, fixers and report rendering.
The repository has about 446 stars and released v7.1.0 on 28 October 2026, the same day the server was published to npm as @community-access/accessibility-agents-mcp; commits were landing in November 2026. A public hosted copy with 12 tools runs at letitglow.app/a11y/mcp with no sign-in.
How to install Accessibility Agents
npx -y @community-access/accessibility-agents-mcpIs Accessibility Agents worth it?
The useful distinction in this package is between what is measured and what is judged. The MCP tools are deterministic: run_axe_scan drives axe-core in a local Chromium with every WCAG 2.2 A and AA tag; the Playwright tools check keyboard navigation, rendered contrast and reflow at several widths; run_verapdf_scan validates PDF/UA-1, PDF/UA-2 or PDF/A; check_contrast computes ratios. The Office, PDF, ePub and Markdown scanners and the check_heading_structure, check_link_text and check_form_labels tools are rule checks that read the file's markup by pattern-matching, so they find a missing alt attribute or a skipped heading, not a bad one.
Everything else is model judgement, delivered by the skills: whether alt text is accurate, whether focus moves sensibly when a dialog closes, whether a custom widget follows its keyboard pattern. The alt-text tools make the split plain: extract_document_images hands each picture to your model, which writes the description, and write_document_alt_text only puts it into the file. The edit gate checks that the accessibility lead skill ran before a file is changed, not that its review was right.
The project's own docs are candid about the gap. Its WCAG page says every report states that automated checks find a minority of barriers, and that axe-core 4.14 has a rule for only one new WCAG 2.2 criterion, target-size (2.5.8), while 2.4.11, 2.5.7, 3.2.6, 3.3.7 and 3.3.8 have no axe rule at all. The figures elsewhere are less tidy: one skill says automated tools catch about 30% of issues, another 30 to 40%, and an integration page claims agents catch about 70% during code generation, with no study cited for any of them.
The catch: the server reaches further than a linter. The browser tools load any http or https address you or the model pass, including localhost and pages with third-party scripts, and that page text flows back to the model. fix_document_metadata and fix_document_headings overwrite documents, within your home directory and working directory. The five GLOW tools upload the named file to letitglow.app. HTTP mode binds 127.0.0.1:3100 with no authentication and prints a warning if moved off loopback.
Our take: this is the closest thing to an accessibility specialist sitting beside the agent, and its honesty about limits is a reason to trust it. A clean run means the machine-checkable part passed; the screen reader pass still needs a person.
Where it earns its place
A front-end developer at a 20-person e-commerce company
Catch accessibility regressions in a checkout redesign before review
Point run_axe_scan and run_playwright_keyboard_scan at the local dev server, fix what they report, and let the edit gate force the accessibility lead review on each component change. Free; plan a manual screen reader pass for focus and announcements.
A university disability services office
Check a folder of lecture PDFs and Word handouts before term
batch_scan_documents takes up to 50 files per call, and run_verapdf_scan adds PDF/UA validation with Java 11 and veraPDF installed. Results go to a local .a11y-history folder so the next run shows the change.
A communications officer at a public body
Add alt text to every picture in a PowerPoint deck
On the hosted server, extract_document_images shows each picture to the assistant, which drafts alt text of at most 150 characters; after review, write_document_alt_text returns the deck behind a one-hour download link.
Before you commit
- Setup
- As a plugin: /plugin marketplace add Community-Access/accessibility-agents, then /plugin install in Claude Code (Copilot CLI, Codex, Gemini CLI and VS Code have equivalents). Server alone: npx -y @community-access/accessibility-agents-mcp. Browser scans need playwright, @axe-core/playwright and Chromium; deep PDF checks need Java 11+ and veraPDF. Node.js 22, 24 or 26.
- Integrations
- 39 tools by the README, plus 3 prompts (audit-page, check-component, explain-wcag) and resources for component guidelines and scan profiles. Read tools cover axe, Playwright, documents, Markdown, contrast and audit history; write tools fix metadata and headings, convert PDF forms to HTML, write reports and accessibility statements. Skills cost about 1,100 tokens before use.
- Data and privacy
- No telemetry. Scans run locally; files are read and written only under your home and working directories, with symlinks resolved. Browser tools load the URLs given, including third-party scripts on those pages, and page content is untrusted input to the model. GLOW tools send files to letitglow.app. The hosted server deletes uploads before answering, except alt-text output kept for one hour.
- Cost at scale
- Free. The only costs are your model's tokens and a machine able to run headless Chromium; the hosted server is free with no sign-in.
- Limits
- Documents capped at 100 MB and batches at 50 files; hosted uploads at 20 MB. Codex installs get no edit gate; ChatGPT gets skills only, with no scanners. The new WCAG 2.2 criteria other than target size have no axe rule. About seven open issues.
What’s good
- Real axe-core, Playwright and veraPDF scans, not model guesses
- Covers Office, PDF, ePub and Markdown as well as web UI
- Edit gate stops unreviewed changes to UI files
- No telemetry, account or key; scans run locally
What’s not
- Most WCAG 2.2 additions still need manual testing
- Document and HTML checks are pattern-matching heuristics
- Coverage figures in its docs are inconsistent and unsourced
- HTTP mode has no authentication
Best for
Front-end teams and document producers who want their coding agent to run real accessibility scans and follow WCAG 2.2 AA guidance before changes land.
Accessibility Agents alternatives
Accessibility Agents: common questions
- Is Accessibility Agents worth it?
- A broad accessibility toolkit for coding agents that runs real scanners and is unusually honest that its scanners prove a floor, not conformance. We rate it 4.1 out of 5.
- How much does Accessibility Agents cost?
- Free and open source (MIT). No account or key; the hosted server is free and needs no sign-in.
- What are the downsides of Accessibility Agents?
- Most WCAG 2.2 additions still need manual testing; document and HTML checks are pattern-matching heuristics; coverage figures in its docs are inconsistent and unsourced; hTTP mode has no authentication.
- Who is Accessibility Agents best for?
- Front-end teams and document producers who want their coding agent to run real accessibility scans and follow WCAG 2.2 AA guidance before changes land.
- How much does Accessibility Agents cost for a team or in production?
- Free. The only costs are your model's tokens and a machine able to run headless Chromium; the hosted server is free with no sign-in.
- Is Accessibility Agents safe for company data?
- No telemetry. Scans run locally; files are read and written only under your home and working directories, with symlinks resolved. Browser tools load the URLs given, including third-party scripts on those pages, and page content is untrusted input to the model. GLOW tools send files to letitglow.app. The hosted server deletes uploads before answering, except alt-text output kept for one hour.
- What are the best Accessibility Agents alternatives?
- The closest alternatives we have reviewed are Playwright MCP and Chrome DevTools MCP.
Answers checked Nov 11, 2026.
Make Accessibility Agents? Show this rating on your site
