Live project · self-hosted
Most people give an AI a terminal and hope for the best. I built the opposite: a set of careful tools an assistant can call to change git, build apps, and put them on the public internet — with rules that stop a bad deploy cold.
You are reading one of those deploys. This page was published by the system it describes.
Not a demo slide. A working ops system designed for an AI operator.
Handing an AI full SSH access is powerful and reckless. MCP (Model Context Protocol) is a standard way to give models small, typed abilities — “deploy this app,” “add this domain” — instead of “here is root, good luck.”
If git push fails or a build breaks, the agent stops and reports. It does not invent a workaround. That is the difference between a toy chatbot script and something you would trust near real traffic.
This site is not a mockup of the pipeline. It is a product of it: own repo, own container, public HTTPS, updated from chat through those tools.
Six live tool servers on this host. Each one is a capability, not a buzzword.
Operate the machine: version control, health checks, deploy hooks, and verification.
Register apps, write files, build images, and bring containers online.
Put public HTTPS in front of each app and keep the edge configuration in git.
Create repositories and wire webhooks so a push can trigger a real deploy.
Point domain names at the server so the internet can find the new site.
Reach the host privately when the public internet is the wrong path.
Same path used when this page was updated.
Two apps. Two repos. One control plane.
Software engineer in Phoenix. By day I work on mainframe credit decisioning at Wells Fargo (COBOL, CICS, high-stakes releases). After hours I build systems where an AI can operate production the way a careful engineer would — structured tools, clear ownership, and an explicit stop when something is unhealthy.