Every AI coding assistant I tried was locked to one machine or one cloud. I needed something different: an IDE where I could wire any AI model to any machine on my network — local Ollama on the GPU box, Claude API from the laptop, streamed to my phone while I'm on the couch. So I built one.

KYLIDESCOPE is a skeuomorphic rack-and-cable IDE where you visually patch AI models across devices. Click a jack on your Mac, click a jack on Claude Opus, and a cable appears connecting them. The wire-routing engine resolves the chain — composing system prompts, tools, and safety guards from whatever processing modules you've patched in between.

But the the moment it went from pipe dream - PIPED REALITY was when, after feeling insecure in a figurative an literal sense, I was surprised by the ease with which I found Tailscale, signed up and minutes later it just worked.

Tailscale is the nervous system that makes it all work. Every machine on my network has a stable 100.x.x.x address. The relay server is just Express + SSE on a Tailscale IP. My Android phone connects to one URL and gets full agentic AI streaming — tool execution, file browsing, session persistence — with zero cloud infrastructure.

And then 4-line bash script called "Insomniac Mode" made me feel like I really could bet the empire on it. Before a critical client demo, I discovered that when my Mac sleeps, the Tailscale mesh drops and the mobile relay dies. Instead of brute-forcing the machine awake with caffeinate, I wrote a Sleepwatcher wake hook that force-pings the Tailscale mesh node after every wake. The tunnel re-establishes in under 5 seconds. The Mac sleeps normally, saves power — but always wakes up ready.

This talk is a live demo and a solo dev story. I'll show the wire-patching interface, stream AI responses from desktop to phone, and (if the demo gods are kind) put the Mac to sleep and wake it back up on stage. No DevOps team. No security engineering background. Just Tailscale, some LaunchAgents, a problem to solve, and the sheer audacity.