Accessibility MCP server for ChatGPT, Claude, Gemini and Cursor

AI assistants already write a large part of the code of many sites. The Pandus MCP server gives them what they lack for accessibility: what is actually wrong on your pages, why it matters for your customers and how to fix it on your platform. The assistant reads the problem, finds the template that renders it in your project, fixes the cause and starts a new check to confirm it.

It works with any assistant that supports remote MCP servers, not with one vendor. Results come from an automated check, not a full audit, and not legal advice.

What the assistant can do

Two ready prompts come with it: fix the most important problems of a site in your codebase, and write an update for the site owner in their language.

Connect your AI assistant

The server address is https://pandus.online/mcp (streamable HTTP). Without a token it serves the free rule reference. For your own sites, create an API token in Pandus under API and webhooks: a read-only token is enough to read; add "write" to let the assistant start checks. Then add the server to your assistant.

Claude (web and desktop)

In Claude, open Customize → Connectors, choose Add custom connector, enter the address and, for your own sites, add the request header Authorization: Bearer YOUR_TOKEN. On Team and Enterprise plans an owner adds the connector in the organization settings.

Claude Code

claude mcp add --transport http pandus https://pandus.online/mcp --header "Authorization: Bearer YOUR_TOKEN"

ChatGPT

On paid plans, turn on developer mode in the settings for apps and connectors, then create a connector with the address https://pandus.online/mcp. Without authentication you get the free rule reference. To read your own sites, use token (bearer) authentication with your API token where your plan offers it. On Business and Enterprise workspaces an admin may need to allow custom connectors first.

Gemini CLI

gemini mcp add --transport http pandus https://pandus.online/mcp --header "Authorization: Bearer YOUR_TOKEN"

Or in ~/.gemini/settings.json:

{
    "mcpServers": {
        "pandus": {
            "httpUrl": "https://pandus.online/mcp",
            "headers": { "Authorization": "Bearer YOUR_TOKEN" }
        }
    }
}

Cursor

In ~/.cursor/mcp.json or .cursor/mcp.json in the project:

{
    "mcpServers": {
        "pandus": {
            "url": "https://pandus.online/mcp",
            "headers": { "Authorization": "Bearer YOUR_TOKEN" }
        }
    }
}

VS Code

In .vscode/mcp.json:

{
    "servers": {
        "pandus": {
            "type": "http",
            "url": "https://pandus.online/mcp",
            "headers": { "Authorization": "Bearer YOUR_TOKEN" }
        }
    }
}

Perplexity and other assistants

Perplexity (Pro, Max and Enterprise) adds remote MCP servers under Settings → Connectors as a custom remote connector. Windsurf, Cline and other MCP clients take the same address; where they ask for authentication, send your token as Authorization: Bearer YOUR_TOKEN.

Keep tokens out of shared repositories: use your assistant's user settings or an environment variable.

What to ask

• "Use Pandus: what are the three most important accessibility problems on shop.example, and where are they in this repository?"
• "Fix the image-alt problem Pandus found on the product pages, then start a new check."
• "Explain WCAG 2.4.7 and check our button styles for it."
• "Compare the last two Pandus checks and write a short update for the client in German."

Pandus and open-source axe MCP servers

Both kinds can run side by side: an open-source server for local tests while you develop, Pandus for the whole site and the team.

Security and data

• Only your team's data. A token sees the sites of its own team, as in the API, and stops working when the plan no longer includes the API.
• Untrusted HTML. Code snippets from your pages are shortened and marked as untrusted, so text hidden in a page cannot give the assistant orders.
• No visitor data. Barriers real visitors met come as totals; session recordings are never sent.
• Limits and history. Requests are limited per token, and every call over your data is written to your team's activity log.

Plans

The rule reference is free for everyone. Your sites, "what to fix", reports and new checks over MCP come with the Pro and Agency plans, the same plans that include the API. See pricing.