Articles, insights, and updates from the Community Access team on AI accessibility, open source development, and assistive technology.
A deep dive into how AI accessibility agents can audit university documents and web properties at scale, covering the document accessibility wizard, web accessibility wizard, and CI/CD integration.
Read articleHigher education institutions are among the largest publishers of digital content in the world. A mid-sized university might maintain hundreds of web pages, publish thousands of PDFs each semester, distribute faculty syllabi in Word, share research in PowerPoint, and post financial aid data in Excel — all of it subject to Section 508, the ADA, and WCAG 2.1 AA.
The tooling has not kept up. Manual audits are slow, expensive, and inconsistent. Automated scanners catch only the most obvious violations. And now that faculty and staff are generating content with AI coding assistants, inaccessible output is accelerating: screen reader traps, missing alt text, presentations without reading order, PDFs with no tagging structure.
Community Access is a suite of specialized AI agents built to close that gap — on both the web and document sides — with no scanner or subscription required.
Most AI accessibility tooling works by giving a language model a long list of WCAG rules and hoping it remembers them. The problem: studies of production AI coding sessions show that skill activation rates without explicit agent enforcement hover around 20%. The model forgets, deprioritizes, or skips accessibility context entirely under token pressure.
Community Access takes a different approach. Each specialist runs in its own dedicated context window with a system prompt that is the accessibility domain. The rules are not suggestions layered on top of a general-purpose assistant — they are the agent's entire identity.
| Platform | How Agents Work |
|---|---|
| Claude Code | Agents invoked via /agent-name for accessibility evaluation on demand |
| GitHub Copilot | Agents invoked via @agent-name in Chat, guided by workspace instructions |
| Claude Desktop | MCP extension with 11 static analysis tools and 6 guided audit prompts |
The document-accessibility-wizard is a seven-phase, guided audit agent for DOCX, XLSX, PPTX, and PDF files. It is purpose-built for the kind of content universities produce every day: syllabi, research reports, budget spreadsheets, presentations, and official publications.
| Phase | What Happens |
|---|---|
| 0 — Discovery | Sets scope, format preferences, and scan profile (strict / moderate / minimal) |
| 1 — File Discovery | Recursively finds all documents, shows inventory by type and count |
| 2 — Scanning | Runs per-file scans with severity scoring and confidence levels |
| 3 — Cross-Document Analysis | Identifies patterns, systemic issues, and template-originated problems |
| 4 — Report Generation | Produces a prioritized report with a metadata dashboard |
| 5 — Follow-Up | Offers remediation scripts, VPAT/ACR export, and template guidance |
| 6 — CI/CD Integration | Generates pipeline configs for automated scanning on commit or publish |
The agent applies 46 built-in rules for Office documents and 56 built-in rules for PDFs, mapped to WCAG 2.1/2.2 success criteria.
/Alt attributeTH scope headersFor departments that publish weekly, the agent supports delta scanning — it uses git diff to identify only files changed since the last commit, scans those, and merges findings with the previous baseline.
Many university accessibility problems originate from shared templates: the university letterhead, the standard syllabus format, the official slide deck. The agent detects when the same violation appears across multiple files and flags it as a template-originated issue, so the fix is applied once to the source.
For departments that need to demonstrate compliance, the agent generates a VPAT 2.5 / ACR compliance report by mapping all findings to WCAG success criteria, suitable for procurement processes or ADA coordinator review.
| Phase | Domain | Specialist Invoked |
|---|---|---|
| 1 | Project discovery and scope | — |
| 2 | Document structure and semantics | alt-text-headings |
| 3 | Keyboard navigation and focus | keyboard-navigator |
| 4 | Forms and input accessibility | forms-specialist |
| 5 | Color and visual accessibility | contrast-master |
| 6 | Dynamic content and live regions | live-region-controller |
| 7 | ARIA usage review | aria-specialist |
| 8 | Data tables and grids | tables-data-specialist |
| 9 | Link text and navigation | link-checker |
| 10 | Document accessibility | word/excel/powerpoint/pdf-accessibility |
| 11 | Testing strategy and tools | testing-coach |
outline: none with no custom replacement<th scope> markup| Agent | What It Enforces |
|---|---|
| aria-specialist | ARIA roles, states, properties, widget patterns |
| modal-specialist | Focus trapping in dialogs, drawers, and overlays |
| contrast-master | Text and UI component contrast ratios, dark mode |
| keyboard-navigator | Tab order, skip links, SPA focus management |
| live-region-controller | aria-live regions, toasts, loading states |
| forms-specialist | Label/input association, error messages, validation |
| alt-text-headings | Alt text quality, heading hierarchy, landmarks |
| tables-data-specialist | Table headers, captions, sortable column announcements |
| link-checker | "Click here" detection, new-tab warnings, duplicate links |
| testing-coach | Screen reader testing, keyboard testing, axe-core CI setup |
| Agent | What It Enforces |
|---|---|
| word-accessibility | DOCX headings, alt text, language, table structure |
| excel-accessibility | XLSX sheet names, chart alt text, merged cells, headers |
| powerpoint-accessibility | PPTX slide titles, reading order, decorative image marking |
| pdf-accessibility | PDF/UA tagging, Matterhorn Protocol, form labeling |
The testing-coach agent can generate a GitHub Actions workflow that installs @axe-core/cli, scans a list of URLs against WCAG 2.1 AA tags, outputs a SARIF 2.1.0 report, and uploads findings to the GitHub Security tab. Violations block the merge.
The document wizard generates a CI workflow that detects changed documents in the diff, runs the appropriate scanner, and fails the build on Critical or Major findings.
| Profile | What Gets Flagged |
|---|---|
| strict | All rules, all severities (Critical, Major, Minor) |
| moderate | All rules, errors and warnings only |
| minimal | Errors only — suitable for CI gates |
macOS / Linux:
curl -fsSL https://raw.githubusercontent.com/community-access/accessibility-agents/main/install.sh | bashWindows (PowerShell):
irm https://raw.githubusercontent.com/community-access/accessibility-agents/main/install.ps1 | iexThe installer detects which AI platforms are present and installs the appropriate agent files. It also sets up auto-updates.
| Platform | Location | What |
|---|---|---|
| Claude Code | ~/.claude/agents/ | 22 agent files |
| GitHub Copilot | .github/agents/ | 22 agent files + copilot-instructions.md |
| Claude Desktop | System extension path | MCP extension with 11 tools + 6 prompts |
| Capability | Document Wizard | Web Wizard |
|---|---|---|
| File formats | DOCX, XLSX, PPTX, PDF | HTML, JSX, TSX, Vue, Svelte |
| Rule count | 46 Office + 56 PDF | All WCAG 2.1/2.2 AA criteria |
| Phase-based guidance | 7 phases | 11 phases |
| Delta scanning | Yes (git diff) | Via axe-core CI |
| Cross-file pattern detection | Yes (template analysis) | Yes (cross-page analyzer) |
| VPAT / ACR export | Yes | Via accessibility-wizard report |
| CI/CD integration | Pipeline configs | GitHub Actions + SARIF output |
| One-click prompts | 9 prompts | 5 prompts |
Repository: github.com/community-access/accessibility-agents
License: MIT