Your site should be an MCP server
This site is an MCP server. Point your client at adrianpatzi.com/mcp and you can ask
Claude what I’ve written, what I’m working on, and what my background is, without reading
a single page of it.
I build MCP servers for a living. It felt stupid to have a personal site that described that instead of being it.
The argument
A portfolio tells you I can do the work. A working endpoint lets you check. One of those is a claim and the other is evidence, and they cost about the same to build.
The usual version of this is a chatbot widget in the corner of the page. I don’t want that. A widget is a demo I control — my model, my prompt, my framing, my rate limit. An MCP endpoint is different: you connect it to your own client, with your own context, and ask whatever you want. If it’s bad, you’ll notice immediately. That’s the point.
What it actually exposes
Tools for the writing (list, read, search), tools for the background, one for the guestbook, and exactly one write: you can leave a note without opening the site.
Nothing that mutates anything I’d miss. The scariest thing it can do is add a card to a wall that I moderate.
The boring part is the interesting part
The 2026-07-28 revision of the spec removed protocol-level sessions and the initialize handshake. Every request is self-contained now.
So most of the work on this endpoint was deleting things. No session map, no session expiry, no reconnect path, no Redis to hold state between requests that no longer need state held. The handler is one route file and it is shorter than the code that draws the pixels behind this page.
That is the honest shape of most of this work. The interesting version of a system is usually the one where you removed the part you were proud of.
Try it
Add it to Claude as a custom connector, or in Claude Code:
claude mcp add --transport http adrian https://adrianpatzi.com/mcpThen ask it something I haven’t anticipated. If it answers badly, that’s useful to me — there’s a tool for reporting that too.