AI School MCP server

Gives any MCP client - Claude Code, Claude Desktop, or anything else speaking the protocol - the ability to search and read the free AI School curriculum: 550+ tracks covering AI engineering, governance, security, and applied AI by profession.

The point is not "an AI that knows about AI". It is that when an assistant answers a question about, say, chunking strategy or the EU AI Act, it can point at a specific lesson you can go and read, instead of producing a fluent paragraph you have no way to check.

Install

One command, nothing to clone. Node 20+ is the only requirement: no API key, no account, no server to run.

Claude Code

claude mcp add ai-school -- npx -y ai-school-mcp

Claude Desktop - add to claude_desktop_config.json:

{
  "mcpServers": {
    "ai-school": {
      "command": "npx",
      "args": ["-y", "ai-school-mcp"]
    }
  }
}

Pin a version if you would rather not track the latest release:

claude mcp add ai-school -- npx -y ai-school-mcp@1.0.1

Hosted, nothing to install. The same server also runs at https://lillytechsystems.com/mcp over Streamable HTTP, and is listed in the official MCP Registry as com.lillytechsystems/ai-school:

claude mcp add --transport http ai-school https://lillytechsystems.com/mcp

From a clone

Faster to start, and the right setup for contributing or for pointing at a local copy of the site:

git clone https://github.com/Lilly-Tech-Collab/ai-school-mcp.git
cd ai-school-mcp
npm install
npm test          # smoke-tests the handshake, tools and a live API call

claude mcp add ai-school -- node "$PWD/server.js"

Set AI_SCHOOL_SITE to point the server at a different host; it defaults to https://lillytechsystems.com.

The server reads the public static curriculum API at https://lillytechsystems.com/ai-school/api/v1/, which is a set of JSON files on the same host as the site. Nothing is authenticated and nothing is written.

Tools

Resources

Notes

• read_lesson only fetches lillytechsystems.com. A "read this URL" tool that accepts arbitrary hosts is an SSRF primitive handed to whoever wrote the prompt, so the host check is deliberate and should stay.
• Search is title-only and deliberately simple. The full-text index is 4 MB and belongs in the browser, not in an MCP round trip; if title matching does not find it, list_tracks then get_track usually will.
• The catalogue is cached per process. A long-running client fetches it once. Restart the server to pick up newly published lessons.
• Content licence. Lessons are free to read, link to and quote with attribution. Republishing whole lessons is not permitted. The metadata (titles, descriptions, URLs) may be used freely with attribution. See aischool://catalogue.

Local development

npm run inspect     # opens the MCP Inspector against this server

The server writes only to stderr; stdout is the protocol channel.