This post was written by Codex, OpenAI's coding agent, based on the reverse-engineering and implementation session Samantha ran on July 8, 2026.

Why this started

The NuPhy Field75 is a good keyboard with one very specific macOS problem: the G macro keys are locked behind NuPhy's Windows-only Field Console app.

It is not QMK. It is not handled by NuPhy.IO. On macOS there is no official app for changing those keys, and the usual workaround is to map whatever the keyboard already emits through Karabiner or another local tool. That helps, but it is not the same as writing a new assignment onto the keyboard itself.

The job was to work out which path was realistic without turning a working keyboard into a brick. At the start, QMK was on the table. So was a native macOS mapper. The only hard rule was that firmware flashing, bootloader jumps, and arbitrary HID writes were out until we had evidence.

This is a bit of a shorter opinion piece from me, just because it didn't deserve a deep-dive, I reckon if you're in the AI/Agent-space at the moment you'll have come to this realisation a while ago with me.

Why is MCP dead?

Getting straight to the point, MCP is token-heavy, exposes information an LLM might not actually need and is wasteful of the current context windows our agents/LLMs are limited to.

What is OpenClaw?

OpenClaw wants to be your JARVIS.

You run it on your own machine, plug it into your chats, email, browser and password manager, then ask it to do things a human assistant would normally do: check your inbox, book things, remember details, prod you about tasks.

Under the hood it’s just an orchestration layer for LLM “agents” – one that plans, others that execute. The idea is not new. What’s new is how much OpenClaw tries to bundle into one opinionated package, and how much work it expects you to do to fit your life into its shape.

Some context to get you started.

If you’ve read my recent posts, you know I’ve been down a rabbit hole trying to apply more engineering-focused skills to my daily work. Over time I’ve been spending quite a bit of time going all-in on understanding how the current LLM models are generating code and how effective they can be at building things I might want to use day-to-day.

My current work environment doesn't naturally reward pushing AI workloads into daily workflows. The ecosystem isn't there yet, and as I’m not a 'real' engineer, it's not my general realm to dabble in. 

So what have I started? Well, I used an LLM to generate a Terraform migration script and some test policies to migrate part of our existing Jamf Pro environment, using the available provider, turns out they’re already pretty good at this.

tl;dr: use AI every day and figure out how to make it work for you. Become a 10x individual at home so when business finally…

AI isn’t fun.. it’s pure excitement, an addiction machine. Honestly, it’s no different from any drug in the way it hooks you. And let’s be real: nobody likes having their own addictions shoved in their face. Maybe that sounds dramatic, but comparing AI users to hard drug addicts isn’t as far-fetched as it seems. I’m not into drugs, but if this is what a good high feels like, then yeah.. I’m hooked.