Typed tools, not a root shell
The assistant does not get open-ended access. It gets MCP servers — Model Context Protocol endpoints with clear jobs: write a file, commit, deploy, attach HTTPS. Power with boundaries.
Live · in production
A full AI ops platform on a real server.
kserpa is a control plane I designed and run myself. An AI assistant can ship websites, manage DNS, and operate the host through six dedicated tool servers — with the same discipline I use on high-stakes enterprise releases.
This page is not a mockup. It is a product of the platform: registered, built, and published through those tools.
What makes it strong
Built like production software — not a weekend script glued to a chatbot.
Hard stops when shipping fails
A failed git push or a broken build ends the run. The agent reports the failure and stops. That is intentional. Reliable systems refuse to paper over bad deploys.
Real products, real isolation
Every app owns its GitHub repo, its container, and its hostname. Lab experiments stay separate from production. The architecture enforces that split.
Six live tool servers
The assistant calls these like APIs. Together they cover the full path from change to public site.
host-control
Server operations: git, health, deploy hooks, and verification probes.
docker-apps
Register products, write source, build images, and start containers.
caddy-proxy
Automatic HTTPS and routing so each app is reachable on its own domain.
github-ops
Repositories and webhooks so a push can drive a real deploy.
hostinger-dns
DNS records so the world can resolve the new hostname.
wireguard-vpn
Private access to the host when public entry is the wrong path.
From request to live site
Every public change follows this path — including updates to this page.
- 1. AskDescribe the change in chat.
- 2. ToolsMCP servers execute the work.
- 3. GitCommit lands on main.
- 4. BuildContainer image is rebuilt.
- 5. ServeHTTPS serves the new version.
Running now
Two products on the same control plane. Click through — they are live.
