Original Summary

Over the past few months I've automated a lot of my work and everyday tasks with AI agents. A laptop that sleeps doesn't work for that. The agents need a machine that's always on, where each job has its own space and one runaway process can't take down the rest. I wanted more than one hosted bot. I wanted several agents, scheduled jobs and test environments running side by side on a machine I control. I also think compute and memory will keep getting more expensive while everyone builds with AI. Hetzner already raised its cloud prices in June: its 8-core, 32 GB dedicated server went from €62.49 to €138.49 a month. After comparing providers, a Hostinger KVM 8 (8 vCPUs, 32 GB of RAM, 400 GB of NVMe) was the best deal I found if you commit for 24 months. The US price is $25.99 a month before tax, renewing at $49.99 after the two years. Prices differ by country: across the 54 Hostinger locales I checked, the same plan ran from about $22 to $32 a month. Paying in my local currency with a 20% referral discount, mine came to about $18. I have two of them now. The catch is that the CPU is shared and metered in credits. A KVM 8 can use all 8 cores while it has credit, and it earns credit back at 4.8 cores' worth. When the credit runs out, the whole VPS drops to 30%, which is 2.4 cores, and it stays there as long as your work keeps pushing. Before I fixed this, Hostinger limited my box 13 times in 10 days, and each time everything on it ran at 30%. So I wrote Hostinger Governor, an open-source tool that budgets the credits. The box gets every core while there's credit, then Governor eases it down to the 4.8-core baseline before the credit runs out. The cores still aren't dedicated, but you keep the steady baseline plus the bursts. The box I'm rebuilding now is designed like this: Nothing is set up by hand. A repo of Ansible code rebuilds the server from an empty VM in about two hours, and the box applies each change itself once it passes checks. No open ports. Everything connects outward, and I reach the box over Tailscale. Each agent or service runs as its own Linux user, so they can't read each other's files. The agents report to me in a private Telegram group, one topic per project. Five kinds of action always wait for my tap: spending money, deleting production data, security changes, messaging all users at once, and sending personal data somewhere new. /freeze stops every agent at once. No agent holds an API key. Keys live in small broker services that check each request, act on it and keep an undo record. The CPU is split into tiers. The always-on agents get a small fixed share, and builds, CI and batch jobs share the rest under Governor's ceiling. When something breaks, the box restarts or rolls back on its own, and an outside service alerts me if it goes silent. You don't need all of this to start. One box, a Linux user per project, notifications on your phone and something watching the CPU credits

  • 情报分类：工作与职业机会
  • 分类依据：内容涉及招聘、求职或职业发展
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/10/15 00:38:05