What they did
Kaliada wanted coding agents that keep working after the laptop is shut, so rented a Hetzner CX33 (4 vCPUs, 8 GB RAM, about 12 euros a month), put it behind Tailscale, and ran the Orca server on it so the same agent sessions open from any device. Work starts as a Linear task with screenshots, or as an OpenSpec spec for bigger changes. An architect agent, mostly Codex, reads the task, splits it, and hands pieces to up to five subagents, each in its own Git worktree. Subagents move and comment on Linear tickets, open pull requests that a reviewer agent can bounce, and merges trigger CI gates. Hermes Agent runs on the same machine; Kaliada chats with it on Telegram to install tools, fix the server, create new projects from a template repo on GitHub, and send a daily AI news digest.
How it went
Kaliada has run it for four months, keeping one to five projects going at once; beyond that the person, not the server, becomes the bottleneck. One test pushed to more than 30 agents at the same time on a .NET and PostgreSQL project, the 8 GB of RAM ran out and everything died; one fresh agent then went through the abandoned worktrees and merged them back into a working state. Setup for each project is still partly manual, and the author says the code arrives faster than it can be reviewed.
Worth knowing
Budget memory per agent, especially when projects need Docker. Worktrees and containers do not stop an agent from touching the host, so permissions still matter, and the workflow is not private just because the server is: Linear, Telegram, GitHub and the model providers are all outside services. Switching between models when one hits a limit is still rough.