MCPCon: open models, open protocols, open memory
AGNTCon + MCPCon in San Jose was a whirlwind. I made it just in time for the keynotes, which were exciting and very technical, and from there it was two days of talks, booths, and conversations that I'm still sorting through weeks later.
In today's issue I'll cover the people I met and the talks that stuck with me, then a live lesson on MCP security horror stories coming on November 13, and some (big!) book news.
Live on Friday, November 13: Emmanuel Paraskakis and I are walking through real MCP security incidents and vulnerabilities in a Maven Lightning Lesson, MCP Security Horror Stories.
Two Days at MCPCon
The Signing
On Thursday morning I signed copies of AI Agents with MCP at the O'Reilly booth. The books flew out of the booth, I think I signed all available copies in under 5 minutes!
Thank you to Demetrios Brinkmann for inviting me to sign, and to everyone who came by.
The People
Shortly before I left for MCPCon, the founders of MCPJam (a platform for testing and debugging MCP servers), Prathmesh Patel and Marcelo Jimenez Rocabado, had reached out to me because they discovered that I mentioned MCPJam in AI Agents with MCP. I made sure to bring them a copy of the book, which I got to present to them after Prathmesh's excellent talk.
They were incredibly stoked, and I spent a lot of the time between talks hanging out at their booth and meeting their whole team. They're building something special and are excelling at one of the hardest parts: assembling a great t
I also met Angelina Yang of TwoSetAI, and we had a great conversation that was all over the place. We were kicking ourselves afterward because we should've recorded it. We're fixing that with an interview later this month.
One highlight was Tim O'Reilly's keynote, which ended with the slogan: open models, open memory, and open protocols. If we're going to make AI work for all of us, and not just enrich the few, we must insist upon and build these things. I got to chat with him at the O'Reilly booth as well, and it took all I had not to be completely starstruck.
Beyond that, I met a bunch of great people, made a ton of connections, got some cool swag, and did a little adult trick-or-treating.
The Talks That I'm Still Thinking About
The talks were valuable across the board, but two stayed with me.
Miriah Peterson's (founder of Haikei Labs) talk inspired me to spend more time with local deployments of LLMs for coding. Something I thought was far out of reach of my embarrassingly meager hardware setup (I have limited space, and it's all taken up by my vinyl and Warhammer collections) was much more doable than I ever thought.
Austin Parker gave a really nice observability talk on using OpenTelemetry spans to find agents' "desire paths" - actions that agents prefer to take - in order to either find failure points or areas to improve. Austin also wrote one of the praise quotes for the book, so he gets thanked twice.
The Livestream
On Friday I sat down with Demetrios Brinkmann for a live interview on the Agentic AI Foundation's livestream all about the book: what it covers, where to find it, and what it's like to write about a protocol that keeps changing under your feet (or asciidoc). If you missed it, this YouTube link starts at the interview. See if you can spot the mistake I made in the opening seconds of the interview 🫠
I learned a ton and came home really inspired. Thank you to everyone at the Agentic AI Foundation who put it together. Conferences are good, and I need to go to more of them.
AI Agents with MCP Is Out!
Okay, the big news is that AI Agents with MCP is here! The ebook came out last month, and the dead tree version went on sale right before I left for MCPCon: print copies hit Amazon on October 20 and everywhere else on October 27. It covers the protocol's architecture and the Python SDK, including the July 2026 changes. If you couldn't make it to the signing, order links are on the book page.
What are you waiting for?
MCP Security Horror Stories: AI Agents vs. Your Product
On Friday, November 13 at 12:30 p.m. ET, Emmanuel Paraskakis (three-time VP of Product, who led Apiary and Swagger and now runs Level 250) and I are hosting a free Maven Lightning Lesson: MCP Security Horror Stories: AI Agents vs. Your Product. We'll walk through real MCP security incidents and, for each one, what happened, why it worked, and the fix.
One of them happened last week: on October 29, Cursor's extension marketplace served tampered versions of the Claude Code extension that added an attacker's SSH key to developers' machines, delivered through auto-update. It's an hour, it's free, and you'll leave with a checklist to run on Monday. Bring your questions and your own horror stories.
This lightning lesson is my first foray into teaching on Maven. Follow me there for updates on upcoming lightning lessons, courses, and more.
Quick Signals
- Miriah's talk is a good excuse to revisit Hugging Face's State of Open Models: Summer 2026 report. Models under 1B parameters account for 83% of all-time downloads, and everything above 100B accounts for 1%. Their own framing, "attention is not adoption," is worth keeping in mind the next time a giant open-weight release tops the leaderboards.
- If you want to try local agentic work yourself, Meta's Muse Glimmer is a 30B open model under Apache 2.0, built for reliable function calling over long workflows, and it runs on a single-GPU Mac or PC in under 20GB quantized.
See you space cowboy...
Kyle
