Use the scanner from an AI tool

The MCP server makes a website scan available to a compatible AI application. MCP means Model Context Protocol: a way for an application to call tools.

This integration runs Agent Lighthouse. It does not make the scanned website support MCP or guarantee that another AI system can use it.

Connect the server

You need Node.js, npm, and a client that supports launching local MCP servers. Add this server entry to your client’s MCP configuration:

{
  "mcpServers": {
    "agent-lighthouse": {
      "command": "npx",
      "args": ["-y", "@forkpoint/agent-lighthouse-mcp"]
    }
  }
}

Your client determines where this file lives and when to reload it. Keep any other server entries you already use.

Ask for a website check

Ask your client to scan a public website address with Agent Lighthouse. Review the findings and suggested fixes. Do not treat the score as proof that every AI system can use that website.

If the tool does not appear, check the client’s server logs and confirm it can run npx. To check the scanner separately from your client, run the command in the quickstart.

Ask what the server checks

Takes no arguments and runs no scan. It returns the installed version’s getRegistrySummary() as JSON: the core version, the number of audits and categories, the count per tier, the registry mass, and the same per category.

Prefer to call it from code?

Use the SDK guide for a Node.js or TypeScript application. You do not need an MCP client for that route.

Page purpose controls

The audit_website tool accepts optional pageType and pages fields:

{
  "url": "https://example.com/guide",
  "pageType": "article",
  "pages": [
    { "url": "https://example.com/products/widget", "pageType": "product" }
  ]
}

Declare a type only when it matches the page’s purpose. Supported types are homepage, category, product, article and general. The retired names unknown and content are refused; use general. Omit the type for detection; type-specific findings then stay advisory. The scan shares the SDK’s validation and evidence rules.

The response includes pageScope when the scan records scope data. It contains page classification, fetch attempts, and separate result populations with their selected, input, and unread URLs. Advisory results do not change the score. Each audit still counts once. The programmatic auditWebsite(url, { pageType, pages }) helper supports the same declarations.

Moving a 6.x integration? See the migration notes.