Accessibility Agents

AI and automated tools are not perfect. They miss things, make mistakes, and cannot replace testing with real screen readers and assistive technology. Always verify with VoiceOver, NVDA, JAWS, and keyboard-only navigation. This tooling is a helpful starting point, not a substitute for real accessibility testing.

A community-driven open-source project automating accessibility, efficiency, and productivity through AI-based agents, skills, custom instructions, and prompts.

A sincere thanks goes out to Taylor Arndt and Jeff Bishop for leading the charge in building this community project. It started because LLMs consistently forget accessibility - skills get ignored, instructions drift out of context, ARIA gets misused, focus management gets skipped, color contrast fails silently. They got tired of fighting it and built an agent team that will not let it slide. Now we want to make more magic together.

We want more contributors! If you care about making software accessible to blind and low vision users, please consider submitting a PR. Every improvement to these agents helps developers ship more inclusive software for the people who need it most.

The Problem

AI coding tools generate inaccessible code by default. They forget ARIA rules, skip keyboard navigation, ignore contrast ratios, and produce modals that trap screen reader users. Even with skills and CLAUDE.md instructions, accessibility context gets deprioritized or dropped entirely.

The Solution

Accessibility Agents provides seventy-nine specialized agents across eight teams and five platforms:

Web Accessibility team - agents that enforce WCAG AA standards for web code, including i18n/RTL, WCAG AAA coverage, data visualization, email, media, web components, and performance accessibility
Document Accessibility team - agents for Office (DOCX, XLSX, PPTX), PDF, EPUB, and Markdown accessibility scanning and remediation
GitHub Workflow team - twenty agents that manage repositories, triage issues, review PRs, manage project boards, CI/CD workflows, security alerts, releases, notifications, and wiki pages
Developer Tools team - agents for Python, wxPython, desktop accessibility, NVDA addon development, and accessibility tool building
Education & Standards - screen reader simulation, WCAG 3.0 preview, WCAG AAA auditing
Cross-cutting - CI/CD accessibility pipelines, compliance mapping, accessibility statements, regression detection, and orchestrators that route work across teams

All agents run on:

Claude Code - Agents you invoke directly for accessibility evaluation
GitHub Copilot (VS Code and CLI) - Agents + workspace instructions that ensure accessibility guidance in every conversation
Gemini CLI - Skills-based extension with always-on WCAG AA context via GEMINI.md
Codex CLI - Direct Accessibility Agents skills pack plus optional experimental TOML-based roles for focused accessibility passes
MCP Server - HTTP-based server providing 24 accessibility scanning tools to any MCP-compatible client (Claude Desktop, VS Code, CI/CD pipelines)

System Requirements

⚠️ CRITICAL: To remain current with Accessibility Agents and ensure proper functionality, you must keep all tools updated to their latest versions. New platform capabilities, API changes, accessibility features, and bug fixes directly impact agent behavior.

Required Tools (Latest Versions)

For GitHub Copilot (VS Code):

VS Code: Latest stable release and/or VS Code Insiders (Download)
GitHub Copilot Extension: Latest version from VS Code Marketplace
GitHub Copilot Chat Extension: Latest version from VS Code Marketplace
Node.js: v18.0.0 or higher (for CLI tools like axe-core)

For Claude Code:

Claude Code CLI: Latest version (Installation)
Claude Subscription: Pro, Max, or Team plan

For Gemini CLI:

Gemini CLI: Latest version (Installation)
Google AI Studio API Key: Active key (Get Started)

Operating Systems:

macOS: 10.15 (Catalina) or later
Windows: Windows 10/11 with PowerShell 5.1+ (pre-installed)

Why Version Currency Matters

Platform API Changes - VS Code Copilot, Claude Code, and other platforms add new capabilities (tool use, context windows, model selection) that agents rely on
Accessibility Features - New platform features directly improve agent effectiveness (browser tools, screenshot analysis, DOM inspection)
Bug Fixes - Critical fixes for tool invocation, context handling, and agent orchestration
Security Updates - Important security patches for API access, authentication, and data handling
WCAG Evolution - As standards evolve (WCAG 2.2, 3.0), agents update to reflect current best practices

Keeping Tools Updated

Automatic Updates (Recommended):

# Set up daily auto-updates during installation (recommended)
curl -fsSL https://raw.githubusercontent.com/Community-Access/accessibility-agents/main/install.sh | bash
# Choose "Yes" when prompted for auto-updates

# Or manually enable auto-updates later
bash update.sh --auto

Manual Updates:

# Update Accessibility Agents
cd accessibility-agents
git pull origin main
bash update.sh

# Update VS Code
# Help → Check for Updates (or auto-updates if enabled in settings)

# Update GitHub Copilot Extensions
# Extensions → @installed → Click update icon next to GitHub Copilot extensions

# Update Claude Code CLI
claude code update

# Update Node.js tools
npm update -g @axe-core/cli
npm update -g pa11y

Version Checks:

# Check current versions
code --version                    # VS Code
claude code --version            # Claude Code CLI
node --version                   # Node.js
npm list -g --depth=0            # Global npm packages

Compatibility Note

Accessibility Agents are tested against the latest stable releases of all supported platforms. While older versions may work, we cannot guarantee compatibility or support issues arising from outdated tooling. If you encounter unexpected behavior, update all tools before reporting issues.

VS Code 1.113 Highlights

VS Code 1.113 is currently the most relevant baseline for GitHub Copilot users of this repo. The release added several changes that improve Accessibility Agents workflows directly:

MCP across agent types - MCP servers configured in VS Code now bridge into Copilot CLI and Claude agents, which makes this repo's MCP guidance more consistent across local, CLI, and Claude workflows.
Chat Customizations editor - Chat: Open Chat Customizations gives you one place to inspect and manage instructions, prompt files, agents, skills, MCP servers, and plugins.
Broader Agent Debug coverage - Agent Debug Logs now cover Copilot CLI and Claude agent sessions in addition to local sessions.
Integrated browser improvements - local HTTPS testing with self-signed certificates is easier, and browser tab management is better for accessibility testing workflows.
Nested subagents - VS Code now supports nested subagent delegation. For this repo, that is treated as an optional platform capability, not a default architecture choice. We favor explicit coordinator-worker flows and keep nested subagents disabled by default.

Subagent stance for this repo:

Reward: bounded subagents improve specialization, parallel analysis, and audit structure.
Risk: unrestricted or nested subagents can increase duplicate findings, wrong-agent selection, token cost, and debugging complexity.
Recommendation: use explicit, allowlisted specialist delegation; avoid recursive subagent chains unless a workflow is intentionally designed for them.

For official details, see the VS Code 1.113 release notes: https://code.visualstudio.com/updates/v1_113.

Authoritative Sources and Currency

This project bases platform-specific guidance on official vendor documentation and release notes, not secondary summaries.

Primary references:

VS Code release notes: https://code.visualstudio.com/updates
VS Code Copilot customization docs: https://code.visualstudio.com/docs/copilot/customization/custom-instructions
VS Code custom agents docs: https://code.visualstudio.com/docs/copilot/customization/custom-agents
VS Code prompt files docs: https://code.visualstudio.com/docs/copilot/customization/prompt-files
GitHub Copilot product docs: https://docs.github.com/copilot

Attribution policy:

Platform claims in this repo should cite at least one official source link.
New behavior tied to a specific release should include the release note URL.
When settings keys are documented, link to the official settings/docs page where possible.

Optional Customization

VS Code 1.113 Workflow Tips

If you are using Accessibility Agents in VS Code 1.113, these settings and commands are the most useful starting point:

{
  "chat.useCustomizationsInParentRepositories": true,
  "github.copilot.chat.agentDebugLog.enabled": true,
  "github.copilot.chat.agentDebugLog.fileLogging.enabled": true,
  "chat.imageSupport.enabled": true
}

Run Chat: Open Chat Customizations to inspect loaded instructions, agents, skills, MCP servers, and plugins.
Use /troubleshoot when a customization is not loading or an expected tool is missing.
Keep chat.subagents.allowInvocationsFromSubagents disabled unless you are intentionally experimenting with recursive orchestration patterns.

That last point is a repo recommendation, not a VS Code limitation.

Custom Thinking Phrases (VS Code 1.110+)

VS Code users: Personalize the loading text that appears while agents think with accessibility-themed phrases.

Add to VS Code Settings (settings.json):

{
  "chat.agent.thinking.phrases": {
    "mode": "append",  // Adds to default phrases
    "phrases": [
      "Checking contrast ratios...",
      "Testing with screen readers...",
      "Verifying keyboard navigation...",
      "Reviewing ARIA patterns...",
      "Scanning for accessibility barriers...",
      "Consulting WCAG 2.2..."
    ]
  }
}

Options:

"mode": "append" - Adds your phrases to VS Code's default list (recommended)
"mode": "replace" - Only shows your custom phrases

Why This Matters:

Reinforces accessibility focus during agent work
Reminds team members that accessibility is actively considered
Optional fun enhancement to make wait time more engaging

How to Add:

Open VS Code Settings (Ctrl/Cmd + ,)
Click "Open Settings (JSON)" icon in top-right
Add the chat.agent.thinking.phrases setting
Reload window (Command Palette → "Developer: Reload Window")

Community Contributions:
Have a great accessibility-themed thinking phrase? Submit a PR to add it to our recommended list in CONTRIBUTING.md!

Quick Start

One-liner install

macOS:

curl -fsSL https://raw.githubusercontent.com/Community-Access/accessibility-agents/main/install.sh | bash

The shell installer requires a real bash environment. On Windows, use PowerShell by default; use the shell installer only from Git Bash, WSL, or another environment that provides bash.

Windows (PowerShell):

irm https://raw.githubusercontent.com/Community-Access/accessibility-agents/main/install.ps1 | iex

If both VS Code and VS Code Insiders are installed, the installer copies Copilot assets into both profiles automatically.

Non-interactive install examples:

bash install.sh --project --copilot --yes --no-auto-update --dry-run

.\install.ps1 -Project -Copilot -Yes -NoAutoUpdate -DryRun

Each install, update, and uninstall command now writes a machine-readable summary or plan file by default. PowerShell scripts accept both -SummaryPath ... and --summary ...; uninstall.ps1 also accepts --summary=... for compatibility. Shell scripts continue to use --summary=....

See the full Getting Started Guide for all installation options, manual setup, global vs project install, auto-updates, and platform-specific details.

One-liner uninstall

macOS:

curl -fsSL https://raw.githubusercontent.com/Community-Access/accessibility-agents/main/uninstall.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/Community-Access/accessibility-agents/main/uninstall.ps1 | iex

The uninstaller removes all agents, config sections, assets, extensions, and auto-update tasks across every platform (Claude Code, Copilot, Codex, Gemini). If no manifest file is found, it downloads the repo to build a fallback file list so nothing is missed.

For step-by-step manual removal instructions, see UNINSTALL.md.

Safe installation — your files are never overwritten

The installer is designed to be additive and non-destructive:

Agent files (~/.claude/agents/, .github/agents/) - existing files are skipped, not replaced. A message tells you which agents were skipped so you know what you already have.
Config files (copilot-instructions.md, copilot-review-instructions.md, copilot-commit-message-instructions.md) - our content is wrapped in  markers and merged into your existing file. Your content above and below the markers is always preserved. If the file does not exist, it is created.
Asset directories (skills/, instructions/, prompts/) - copied file-by-file; files that already exist are skipped.
Manifest file (.a11y-agent-manifest) - tracks every file we installed. The update script uses this list to ensure it only touches files we own, never user-created agents. When contributors add new agents to the repo, those files are automatically installed on next update and added to the manifest.

Updates are equally safe - the update script never deletes agent files. If a file is not in the manifest (meaning you created it yourself), it will not be modified or removed.

To reinstall a specific agent from scratch, delete it first and rerun the installer (or update script).

Install from VS Code Marketplace (Coming Soon)

GitHub Copilot users (VS Code): A VS Code extension is in development that will allow one-click install from the Extensions marketplace. Until then, use the one-liner install above or the manual steps in Getting Started.

What the extension will include:

✅ 59 fully-integrated agents in GitHub Copilot (VS Code and CLI)
✅ 18 reusable accessibility skills (WCAG rules, severity scoring, scanning patterns)
✅ 106 custom prompts for web audits, document audits, GitHub workflows, and developer tooling
✅ 6 workspace instructions (automatic WCAG AA enforcement on every chat)
✅ 100% source citation coverage (all agents cite authoritative standards)
✅ Auto-update mechanism (new agents and features arrive automatically)

For other platforms (Claude Code, Gemini, Claude Desktop, Codex), see Getting Started.

The Team

The following agents make up the accessibility enforcement team, each owning one domain.

Agent	Role
accessibility-lead	Orchestrator. Decides which specialists to invoke and runs the final review.
aria-specialist	ARIA roles, states, properties, widget patterns. Enforces the first rule of ARIA.
modal-specialist	Dialogs, drawers, popovers, alerts. Focus trapping, focus return, escape behavior.
contrast-master	Color contrast ratios, dark mode, focus indicators, color independence.
keyboard-navigator	Tab order, focus management, skip links, arrow key patterns, SPA route changes.
live-region-controller	Dynamic content announcements, toasts, loading states, search results.
forms-specialist	Labels, errors, validation, fieldsets, autocomplete, multi-step wizards.
alt-text-headings	Alt text, SVGs, icons, heading hierarchy, landmarks, page titles.
tables-data-specialist	Table markup, scope, caption, headers, sortable columns, ARIA grids.
link-checker	Ambiguous link text, "click here" detection, missing new-tab warnings.
accessibility-wizard	Interactive guided web audit across all eleven accessibility domains.
testing-coach	Screen reader testing, keyboard testing, automated testing guidance.
wcag-guide	WCAG 2.2 criteria in plain language, conformance levels, what changed.
word-accessibility	Microsoft Word (DOCX) document accessibility scanning.
excel-accessibility	Microsoft Excel (XLSX) spreadsheet accessibility scanning.
powerpoint-accessibility	Microsoft PowerPoint (PPTX) presentation accessibility scanning.
office-scan-config	Office scan rule configuration and preset profiles.
pdf-accessibility	PDF conformance per PDF/UA and the Matterhorn Protocol.
pdf-scan-config	PDF scan rule configuration and preset profiles.
document-accessibility-wizard	Guided document audit with cross-document analysis, VPAT export, and CSV export with help links.
markdown-a11y-assistant	Markdown documentation audit — links, alt text, headings, tables, emoji, diagrams, em-dashes, anchors.
text-quality-reviewer	Catches invisible text quality issues: template variables in alt text, code syntax as accessible names, empty labels, duplicate control labels.

Developer Tools Agents

The following agents support Python, wxPython, desktop accessibility, NVDA addon development, and accessibility tool building.

Agent	Role
developer-hub	Orchestrator. Routes development tasks to the right specialist from plain English.
python-specialist	Python debugging, packaging (PyInstaller/Nuitka/cx_Freeze), testing, type checking, async, optimization.
wxpython-specialist	wxPython GUI — sizer layouts, event handling, AUI, custom controls, threading, desktop accessibility.
desktop-a11y-specialist	Platform accessibility APIs (UI Automation, MSAA/IAccessible2, NSAccessibility), screen reader Name/Role/Value/State, focus management.
desktop-a11y-testing-coach	Screen reader testing with NVDA, JAWS, Narrator, and VoiceOver. Automated UIA testing, keyboard-only testing flows.
a11y-tool-builder	Build accessibility scanning tools, rule engines, document parsers, report generators, and audit automation.
nvda-addon-specialist	NVDA screen reader addon development — globalPlugins, appModules, synthDrivers, braille tables, Add-on Store submission, grounded in official NVDA source.

GitHub Workflow Agents

The following agents handle GitHub repository management, triage, and workflow automation.

Agent	Role
github-hub	Orchestrator. Routes GitHub management tasks to the right specialist from plain English.
daily-briefing	Morning overview - open issues, PR queue, CI status, security alerts in one report.
pr-review	PR diff analysis with confidence per finding, delta tracking, and inline comments.
issue-tracker	Issue triage - priority scoring, duplicate detection, action inference, project board sync.
analytics	Repository health scoring (0-100/A-F), velocity metrics, bottleneck detection.
insiders-a11y-tracker	Track accessibility changes in VS Code Insiders and custom repos with WCAG mapping.
repo-admin	Collaborator management, branch protection rules, access audits.
team-manager	Onboarding, offboarding, org team membership, permission management.
contributions-hub	Discussions, community health metrics, first-time contributor insights.
template-builder	Guided wizard for issue/PR/discussion templates - no YAML knowledge required.
repo-manager	Repository scaffolding - labels, CI, CONTRIBUTING, SECURITY, issue templates.

See the Agent Reference Guide for deep dives on every agent, example prompts, behavioral constraints, and instructor-led walkthroughs.

Documentation

Accessibility Docs

The following guides cover web and document accessibility features.

Guide	What It Covers
Getting Started	Installation for Claude Code, Copilot (VS Code and CLI), Gemini CLI, Claude Desktop, and Codex CLI
User Guide	Comprehensive ecosystem guide with per-platform walkthroughs, all agents, skills, prompts, and recipes
Agent Reference	All 80 agents with invocation syntax, examples, and deep dives
MCP Tools	Static analysis tools: heading structure, link text, form labels
axe-core Integration	Runtime scanning, agent workflow, CI/CD setup
VPAT Generation	VPAT 2.5 / ACR compliance report generation
Office Scanning	DOCX, XLSX, PPTX scanning with 46 built-in rules
PDF Scanning	PDF/UA scanning with 56 built-in rules
Scan Configuration	Config files, preset profiles, CI/CD templates
Custom Prompts	Nine pre-built prompts for one-click document workflows
Markdown Accessibility	Four prompts for markdown auditing, quick checks, fix mode, and audit comparison
Configuration	Character budget, troubleshooting
Architecture	Project structure, why agents over skills/MCP, design philosophy
Subagent Architecture	Coordinator-worker patterns, delegation rules, allowlist validation, nested subagent policy
Troubleshooting	MCP server issues, agent configuration, performance, platform-specific debugging

GitHub Workflow Docs

The following guide covers all GitHub workflow agents and their invocation syntax.

Guide	What It Covers
GitHub Workflow Agents	All 10 workflow agents with invocation syntax, examples, and instructor-led walkthroughs

Advanced Guides

The following guides cover advanced configuration, cross-platform handoff, and distribution.

Guide	What It Covers
Cross-Platform Handoff	Seamless handoff between Claude Code and Copilot
Advanced Scanning Patterns	Background scanning, worktree isolation, large libraries
Plugin Packaging	Packaging and distributing agents for different environments
Platform References	External documentation sources with feature-to-source mapping
Experimental Codex Multi-Agent Roles	Optional TOML-based Codex roles for focused accessibility passes
Research Sources	Authoritative sources (W3C APG, WebAIM, WCAG 2.2, Deque) that informed every agent rule

What This Covers

WCAG 2.1 Level AA compliance
WCAG 2.2 Level A and AA criteria (VPAT/ACR generation)
Screen reader compatibility (VoiceOver, NVDA, JAWS)
Keyboard-only navigation
Focus management for SPAs, modals, and dynamic content
Color contrast verification with automated calculation
User preference media queries (prefers-reduced-motion, prefers-contrast, prefers-color-scheme, forced-colors, prefers-reduced-transparency)
Live region implementation for dynamic updates
Semantic HTML enforcement
Static analysis of headings, link text, and form labels
VPAT 2.5 / Accessibility Conformance Report generation
Office document accessibility scanning (DOCX, XLSX, PPTX) with 46 built-in rules
PDF document accessibility scanning per PDF/UA and the Matterhorn Protocol with 56 built-in rules
Markdown documentation accessibility scanning across 9 domains (links, alt text, headings, tables, emoji, diagrams, em-dashes, anchors, plain language)
SARIF 2.1.0 output for CI/CD integration
CSV export with help documentation links for web and document audit findings
Common framework pitfalls (React conditional rendering, Tailwind contrast failures)
NVDA screen reader addon development (globalPlugins, appModules, synthDrivers, braille tables, Add-on Store submission)
Desktop application accessibility (UI Automation, MSAA/IAccessible2, NSAccessibility)

Source Citation Policy

Every agent follows a formal source citation policy. AI giving accessibility advice must be held to a higher standard -- wrong guidance creates real barriers for real people.

No source, no claim. If an agent cannot cite an authoritative source, it explicitly flags the recommendation as experience-based.
Inline citations. Every factual claim includes a link to the source.
Six-tier authority hierarchy. Normative specs (WCAG, ARIA) > Informative guidance > Platform vendor docs > AT vendor docs > Peer-reviewed experts > Government/legal.
Machine-readable source registry. SOURCE_REGISTRY.json maps every agent domain to its designated primary authorities.
Automated freshness checks. A weekly GitHub Actions workflow verifies source URLs are still live and opens issues when documentation drifts.

See CITATION_POLICY.md for the full policy.

Roadmap

See ROADMAP.md for what is planned, in progress, and shipped. Track individual items on the roadmap issues board.

What This Does Not Cover

Mobile native accessibility (iOS/Android). A separate agent team for that is planned.
WCAG AAA compliance (agents target AA as the standard). An AAA agent is planned.

Example Project

The example/ directory contains a deliberately broken web page with 20+ intentional accessibility violations. Use it to practice with the agents and see how they catch real issues. See the example README for details.

Contributing

This project exists because the community shows up. Every feature in v2.5 -- the NVDA Addon Specialist, the Text Quality Reviewer, the source citation policy, the wxPython screen reader documentation -- started as a community conversation. Not a roadmap item. A real person saying "this is what I need."

Whether you are a developer, accessibility specialist, screen reader user, or just someone who cares about inclusive software - there is a place for you here.

Found an agent gap? Open an issue describing what the agent missed or got wrong.
Know a pattern we should catch? Open a PR. Agent files are plain Markdown - no special tooling required.
Building for the blind and low vision community? Your lived experience and domain knowledge are exactly what makes these agents better. We would love your involvement.

See the Contributing Guide for full details, guidelines, and how to get started.

If you find this project useful, please star the repo and watch for releases so you know when updates drop.

Contributors

A sincere thanks to Taylor Arndt and Jeff Bishop for leading the charge, and to every community member who has contributed to making AI coding tools more accessible.

Resources

Web Content Accessibility Guidelines (WCAG) 2.2
WCAG 2.2 Understanding Documents
WAI-ARIA Authoring Practices Guide
WAI-ARIA 1.2 Specification
Deque axe-core Rules
NVDA User Guide
VoiceOver User Guide
WebAIM Contrast Checker
WebAIM Screen Reader User Survey
A11y Project Checklist
Inclusive Components
Research Sources and Attribution - All authoritative sources consulted to build and strengthen agent rules

Swift Agent Team - 9 specialized Swift agents for Claude Code. Swift 6.2 concurrency, Apple Foundation Models, on-device AI, SwiftUI, accessibility, security, testing, and App Store compliance.

License

MIT

About This Project

Accessibility Agents was founded by Taylor Arndt (COO at Techopolis) and Jeff Bishop because accessibility is how they work, not something bolted on at the end. When AI coding tools consistently failed at accessibility, they built the team they wished existed - and opened it to the world.

This is a community project. The more perspectives, lived experiences, and domain knowledge that go into it, the better it serves the blind and low vision community. If you have ideas, open a discussion. If you have fixes, open a PR. Every contribution matters.