The best VPS for OpenClaw: real requirements, not guesswork
What OpenClaw actually needs from a server — measured, not assumed — and how to choose a provider, size, and setup that won't bite you later.
A small instance genuinely suffices: 2 vCPU and 4 GB of RAM runs a personal OpenClaw agent with headroom — we’ve measured the idle agent around 540 MB with a full browser workload proven on exactly that size. Any reputable provider with that tier for $5–25/mo works; OpenClaw publishes official guides for the major ones. What separates a good setup from a regretted one isn’t the provider — it’s backups, a pinned version, and never exposing the Gateway.
Real requirements, measured
Numbers from operating OpenClaw pods rather than from guesswork (mid-2026, current stable line, Docker deployment):
- RAM is the binding constraint. The agent container idles around 540 MB; a 2 GB machine can run a lean agent, and 4 GB gives comfortable headroom for browser work (headless Chromium is the hungry part) plus the OS. One sharp edge from the official docs: building the Docker image needs ~2 GB free or the build gets OOM-killed — on small machines, use the prebuilt images and the constraint disappears.
- CPU barely matters. The heavy thinking happens at your model provider’s end; the agent machine mostly orchestrates. 2 shared vCPUs are fine.
- Disk: 20–40 GB. OS, images, and an agent’s workspace fit easily; media-heavy use grows over time, not explosively.
- Location: near you is unnecessary. You never browse to the pod interactively for daily use — chat rides outbound channels — so pod latency is irrelevant to the experience. Pick on price and jurisdiction preference, not ping.
Choosing a provider
OpenClaw’s docs ship setup guides for the major VPS providers — Hetzner, DigitalOcean, Hostinger, Fly.io, and others — which is a practical shortlist by itself. At the ~4 GB tier the European value providers (Hetzner’s small instances run a few euros a month) undercut the US majors substantially for identical workloads; all of them run this job fine. Three things worth actually checking, because they differ more than compute does:
- Snapshot/backup pricing and restore workflow — you will want automatic backups on (typically ~20% of the instance price), and a restore you’ve tested once.
- A cloud firewall you can configure to expose SSH only.
- Honest capacity — budget tiers at popular providers go in and out of stock; if a size is unavailable, the next tier up costs a few euros more and changes nothing else.
The setup is the actual differentiator
The uncomfortable truth about “best VPS for OpenClaw”: provider choice is maybe a tenth of the outcome. The rest is the setup discipline on whichever box you pick — key-only SSH, the Gateway never on a public interface (double-check this in Docker setups, where the default bind is broader than you’d expect), a pinned version, state volume-mounted and backed up, and openclaw security audit run after setup and every change. The full standard is the safe-hosting checklist; it’s an afternoon to do well and a permanent duty to maintain.
If that list reads like a good weekend, a VPS is a great home for your agent. If it reads like a second job, an Everpod pod is the same architecture — small dedicated machine, pinned image, private gateway, tested backups — operated for you at $29/mo, with your own model key. Same agent either way; the difference is whose checklist it is.