An empty repo became a working voice assistant in one eveningAn empty repo became a working voice assistant in one evening
July 24, 2026 · git init to a live Alexa skill, a Home Assistant hookup, a coding-agent dispatcher, and a web terminal, all in about ninety minutes.Built from nothing to a working voice assistant in one evening: smart-home control, spoken questions, and jobs it runs in the background.
last updated
From empty repo to live skill
This was the founding day for Big Brain. The repo didn't exist at 10:30pm, and by a bit after midnight it was a self-hosted Alexa skill that could turn on lights, answer spoken questions with live web search, and hand off coding jobs to an AI agent with real access to its own codebase. All of it was reached through an outbound-only Cloudflare Tunnel with no ports opened on the home network.
Phase 1-2: infrastructure and Home Assistant
The build went phase by phase, and a couple of them ran as background agents in separate git worktrees so pieces could come together in parallel instead of one at a time. Phase 1 was pure infrastructure: git init, a caffeinate LaunchAgent so the Mac mini stops sleeping, cloudflared installed with a tunnel config template. Phase 2 brought up Home Assistant in Docker with a dependency-free REST client for talking to the Hue bridge, with bridge networking instead of host networking, since host mode turned out not to forward ports on this machine's Docker Desktop.
Phase 3-4: the Alexa backend and job dispatcher
Phase 3 scaffolded skill-backend: the Express server that verifies every Alexa request is cryptographically genuine (certificate chain walked all the way to Node's built-in CA bundle) before routing it anywhere. Phase 4 built task-runner, the piece behind voice-triggered coding jobs: a fixed command map for the fast path, falling through to a real claude -p spawn for anything freeform, with every dispatched job logged to SQLite so its status survives a restart. One fix landed mid-build: the Home Assistant adapter had assumed flat exported functions when the real client turned out to be a class.
Phase 5, and the one loose end
Phase 5, marked optional, shipped anyway: a ttyd web terminal behind the same tunnel, so the Mac mini is reachable from a browser without opening anything new to the internet. The day closed with the tunnel's UUID filled in for real and a status update noting everything was done except one loose end: the Alexa simulator couldn't yet reach the live skill, chalked up to certificate propagation and left to retry later.
This was the day Big Brain was born. At 10:30pm there was no project at all, and by a bit after midnight it was a real, working voice assistant living on an Alexa device, one that could turn smart lights on and off, answer spoken questions using live web search, and even hand off coding tasks to an AI that could edit its own code. All of it ran from a computer at home, reachable securely from the internet without opening up the home network to anything.
The work happened in stages. First came the basic plumbing: keeping the home computer awake, and setting up a secure tunnel so it could be reached from outside without exposing it directly. Next, the smart-light system got wired up, with a small piece of code written to talk to the lights directly.
After that came the core of the assistant: the part that checks every voice request is genuinely coming from Amazon's Alexa service before doing anything with it, and the part that handles voice-triggered coding requests, using a fast path for common commands and falling back to a full AI coding session for anything more open-ended, with every job tracked so it survives a restart. One bug got caught and fixed along the way, where an early assumption about how the smart-light code was structured turned out to be wrong.
A bonus feature also shipped: a web-based terminal, reachable through the same secure tunnel, so the home computer can be managed from a browser without opening anything new to the internet. By the end of the day, everything was working except one loose end: a brand-new voice skill sometimes takes a little while to become reachable after it's first set up, so testing it fully had to wait.