MCP server

Talk to this site from your AI assistant

This site is also a Model Context Protocol (MCP) server. Connect your AI assistant (Claude Code, Claude, Cursor or any MCP client) and it can search and read every post in English and Spanish, follow the RL from scratch series in order, and check the portfolio and its live demos, always citing the original URL.

It is read-only. The server returns text from the posts and does not generate answers of its own. The one exception is asking for a live demo, and I still approve each request by hand, by email.

Endpoint

https://mcp.alexisalulema.com/mcp

Install

Claude Code

claude mcp add --transport http alexisalulema https://mcp.alexisalulema.com/mcp

Claude (web / desktop)

Settings → Connectors → Add custom connector, and paste the endpoint URL.

Any client (JSON config)

{
  "mcpServers": {
    "alexisalulema": {
      "type": "http",
      "url": "https://mcp.alexisalulema.com/mcp"
    }
  }
}

MCP Inspector

npx @modelcontextprotocol/inspector

Run it, choose "Streamable HTTP" and paste the endpoint URL.

Tools

  • search_posts Search the blog (EN/ES) by section, with snippets and links.
  • get_post Read a full post or one section, with its canonical URL.
  • list_posts List posts by language, tag or date.
  • list_tags All tags with how many posts use each.
  • get_series The "RL from scratch" series in reading order.
  • list_projects Portfolio projects and which have a live demo.
  • get_demo_status Whether a live demo is offline, pending or running.
  • request_demo Ask for demo access; I approve every request by email. enabled later

Try asking

  • “Explain PPO clipping using Alexis’s RL from scratch series, citing the sections.”
  • “What did Alexis learn building the Dreaming Agent? Quote the numbers.”
  • “Which of Alexis’s live demos fits someone interested in multi-agent systems?”

Limits and privacy

  • Read calls are rate-limited per client (about 120 per minute). Demo requests are limited to 3 per client and 2 per email per day.
  • Search queries and emails are not stored. Analytics only count which method or tool was called.
  • Search is lexical (BM25) rather than semantic, so if an answer looks off, try rephrasing with the words the post would use.

The content index behind the server is public too: /mcp/index.json