What is a VPS? A plain-language guide for agent owners
A virtual private server is a rented slice of a real computer that stays on when your laptop doesn't. What you actually get for a few euros a month, and why AI agents changed who needs one.
A VPS (virtual private server) is a rented slice of a real computer in a data center. Virtualization software divides one physical machine among many customers, and your slice behaves like a complete computer of your own: your operating system, your administrator access, your own IP address, running 24/7. You rent it by the month for about what two coffees cost: as of mid-2026 a 2-vCPU, 4 GB machine from a European provider is about €6.50/month. It exists the moment you click create, no hardware owned. For years the typical VPS tenant was a website or a hobby project. The new tenant is an AI agent that needs a computer that doesn’t sleep.
“Virtual” and “private,” unpacked
The physical machine in the data center is enormous by laptop standards: dozens of processor cores, hundreds of gigabytes of RAM. A layer called a hypervisor (KVM is the one you’ll see most) carves it into isolated virtual machines, each with its own reserved memory, disk, and share of the processors, each booting its own operating system. Virtual means your computer is one of those slices rather than its own box. Private means the slice is yours alone: unlike old shared web hosting, where many customers’ sites lived in one system as neighbors, your VPS has its own OS install that nobody else logs into. The spectrum runs: shared hosting (a folder on someone’s server), VPS (your own virtual machine on shared hardware), dedicated server (the whole physical box), with price and isolation rising at each step.
What you actually get for the money
Creating a VPS takes about a minute in a provider’s console: pick a location, a size, and an operating system image (Ubuntu Linux is the overwhelming default), and you’re handed an IP address and root (administrator) access. From there it’s genuinely your computer: you connect over SSH, install anything, open or close network ports, reboot it, break it, restore it from a snapshot. Billing is monthly or hourly, and the entry tiers are cheap: small shared-vCPU instances from the big European providers run roughly €4–8/month as of mid-2026, with US providers a few dollars more. Two properties matter more than the specs. It’s always on: no lid to close, no household power cut, batteries and generators behind the rack. And it’s disposable: destroy it and the billing stops; snapshot it and an identical copy can exist somewhere else in minutes. What you don’t get: a screen. A VPS is headless: you drive it over the network, which is the main adjustment for anyone whose mental model of a computer includes a desktop. A VPS is also only one of the machine shapes agents run on now. The VM-and-sandbox map places it against per-second sandboxes and managed pods.
Why AI agents changed who needs one
An agent like OpenClaw or Hermes is a long-running process: it holds memory, watches channels, runs scheduled jobs at 6am, answers from your pocket at midnight. Run it on a laptop and it lives exactly as long as the lid stays open: overnight schedules simply don’t fire on a sleeping machine. A VPS is the cheapest answer to that: a computer that keeps being a computer while yours isn’t. There’s a security dividend too: an agent that executes commands is safer on a machine that isn’t your personal laptop, where the worst it can reach is its own small world. The trade is that the VPS’s upkeep (SSH keys, updates, backups, firewalls) becomes yours; that operations burden is the actual product behind managed agent hosting, where someone else carries it.
The words you’ll meet next
SSH: the encrypted remote-terminal protocol you administer it through; key-based login is the secure norm. Root: the all-powerful administrator account (day-one hygiene is doing routine work as a lesser user). Snapshot: a point-in-time image of the whole disk you can restore or clone. vCPU: a virtual processor core; “shared” vCPUs are the budget tier and are plenty for an agent, which mostly waits on model APIs rather than computing. Firewall: the filter deciding which ports the internet can reach; an agent box should expose almost nothing. Sizing, providers, and prices get their own treatment in our VPS-for-OpenClaw guide.