@kylestratis.com
📊 Data, AI/ML, and Cloud consultant @ Stratis Data Labs 🚉 VPEng @ Sequel Institute 🤓 Startup advisor ✒️ Writing about MCP 🐍 Python ☁️ AWS 🧙♂️ Elixir 🐷Ham radio 🇺🇸↔🇬🇷 [Boston // Athens] Mostly tech things here
Security has been a hot topic in the MCP world since its launch, and this chapter aims to consolidate all the current best practices on the topic in one place.
If you subscribe to O'Reilly's learning platform, you can read it now!
learning.oreilly.com/library/vie...
What I think is the most important chapter of AI Agents with Model Context Protocol (MCP), and most difficult to research and write, is now available in early release. In chapter 7, you'll learn how to test, secure, and distribute your MCP servers.
I have (nearly) a whole chapter about it in my book and I still don't think I did it justice
* nearly every reward schedule is intermittent
I lean towards the latter though because in the literature it’s “intermittent reinforcement” not reward, and IR is a very large class* of reward schedules that have pretty different behavioral effects and consequences. Slot machines employ variable reinforcement, a specific intermittent schedule.
Mostly because it’s so far off from an actual intermittent reward that to call it such is to completely misunderstand either LLMs or Skinnerian behaviorism. Or I guess both is also a possibility.
I initially wrote this up for an internal team as they transitioned to Claude Code from Gemini, it’s more of a ramp-up on both generative coding techniques and Claude-specific ones rather than a head-to-head comparison, but I decided to make a public version and could be helpful:
Oh, and you can preorder it from Amazon and probably other bookstores that you know and love now
I've finished the first draft of chapter 7 of AI Agents with MCP, which you should be able to see on @oreilly.bsky.social's learning platform either later this month or in March.
Before edits, I've written just over 69,000 words, made 22 images, made ~227 physical pages (est), w/ 2 chapters to go
The piece for anyone interested andymasley.substack.com/p/the-ai-wat...
Anthropic has a claude skill for writing MCP servers that works pretty well if you don’t care about tokens. I wrote (am writing) a whole ass MCP book and still preferred to have Claude build one I did to scratch a particular itch of mine
Code has never been worth much of anything by itself, though. patio11’s advice has been true my entire career.
Companies aren’t going to stop needing the product cycle. They might need fewer people whose entire world is the computer, but the practice of making computers do stuff will persist.
i think the biggest issue where i split ethically in a meaningful way with people is that i have long been against intellectual property as a concept. i think IP helps disney, not small artists. a lot of people want to strengthen IP in response to ai stuff and i just don't think that's a good thing
Thanks for reading, Void!
This bout of mania (not really) also got me to write a new issue of The Signal Path, where I talk about the recent explosion in usage patterns for agent coding, memory agents like @void.comind.network, agent orchestration, and goings-on in the MCP world, including chapter 6 of my book going into ER
The image of him fearfully at his door as a disgusting ice pig looms over him is heart rending and enraging.
Now what ATProto blog can I crosspost to?
I'm loving everything the Prefect team is doing for the MCP community
I wrote up an internal guide to Claude Code use, then I cleaned it up and made it public. It's wicked long, but use the TOC and use it more as a reference library than a singular article, and I think it should be really helpful!
Just got everything lined up to release `ed3d-plugins`, the Claude Code harness I use in my day-to-day. While they're yelling about gas towns, I present an e-bike, just don't drive it on the sidewalk.
I'm biased, but I think it's good. I've tried a few similar, and I keep going back.
If you subscribe to O'Reilly's e-learning platform, the raw and unedited manuscript is available for you to spend your weekend with now. Read it here: learning.oreilly.com/library/vie...
It's Early Release Day! AI Agents with MCP just had a major release, with additional details added to all existing chapters, as well as the addition of Chapter 6, which covers advanced techniques for building MCP servers.
I’ve started doing this with simple games for my daughter. It’s really fun and I can quickly spin up simple educational word games at her level based on what she likes.
I don’t think there are many TUI wardialers in training corpuses, yet I was able to get one reasonably together while drunkenly shouting at my phone and doing no testing until I was home and barely cognizant later that night. It would be more complete but was for a themed hackathon
I got inspired by @simonwillison.net's simple HTML tool workflow:
simonwillison.net/2025/Dec/10/...
So I decided to spend a few minutes today to do something similar: tools.kylestratis.com
Claude skills are a good analogy to where this could go, but they’re really only available to code and desktop.
SEPs are frozen for another week but might be worth bringing up. I could see hacking something using a tool for writing and then a prompt for reading and injecting (or even using sampling within the prompt function to better select relevant memories), but it still feels hacky.
Empire of AI is wildly misleading on AI water use - Andy Masley andymasley.substack.com/p/empire-of-ai… (worth reading the details) #AI #water
Just presented this! You can check out the code here:
(note: the layout is screwy and config options don't get shown, but not bad for coding by voice while out and about and not even looking at the results until around midnight). No code was actually touched by me.
github.com/kylestratis/...
The key that so many interviewers get wrong is, of course, the “tightly scoped” part 😅