Where to host OpenClaw so it stays up

A self-hosted agent is only interesting when it is awake. The whole promise — it watches something, it reacts to a message, it does the job overnight — evaporates if it lives on a laptop that spends fourteen hours a day closed. Setup takes an evening; keeping it up is the ongoing thing, which is why “where do I run this” is the question the community actually argues about.

Three shapes, in ascending order of how much of it is your problem to have running at 3am.

An old laptop with the lid-close behaviour changed, a Raspberry Pi, a desktop that never sleeps, a home server. Free if the hardware is sitting there, private in the sense that nothing leaves your house except the model calls, and pleasant to work on because you can put your hands on it.

What you take on is uptime you have never had to think about before: a power cut, a router reboot, a dynamic IP, an OS update that reboots at 4am and does not bring the agent back. None of it is hard. All of it is yours, and you find out it happened by noticing the agent has been quiet.

The common answer, and a good one. A small Linux server that never sleeps, from a few dollars a month, with someone else owning the power and the network.

Size it properly: 2 vCPU and 4 GB of RAM is the practical floor for a comfortable setup, and if your agent drives a real browser, the number jumps — headless Chrome will happily eat several gigabytes on its own, and 8 GB is a more honest starting point for that. Disk goes faster than you think once caches and logs accumulate.

There are free tiers worth knowing about: at least one large cloud provider offers a permanently free ARM instance generous enough to run this. It is a real option, with real caveats — capacity is often unavailable in popular regions, and free tiers get reclaimed on terms you do not control. Fine for a hobby agent, thin ice for something you rely on.

The recurring cost is attention rather than money: restarting after a reboot, updating the agent, rotating credentials, and debugging it at the exact moment you wanted to be doing something else. If you enjoy that, this is genuinely the best-value option on the page. The full shape of it — sessions that survive a dropped connection, systemd, what breaks later — is the same as our Claude Code VPS runbook, which applies almost line for line.

A small industry now exists to run this for you: pick a plan, click deploy, get a private instance and a browser terminal, skip the server administration entirely. It works, it is quick, and for most people it is worth the markup over doing it yourself.

Three questions worth asking before you pay, because the answers vary a lot between providers and rarely appear on the pricing page. Can you leave with your configuration and data, or is it locked in their console? What actually happens to what the agent stores — where does it live, who can read it? And is the model access included, or is the monthly fee purely for the box?

That last question is the one that reorders the comparison. A self-hosted agent is bring-your-own-model: it needs an API key, or a subscription it can drive, and that is a separate bill from the hosting.

So a $4 host is not a $4 setup — it is $4 plus whatever the agent spends thinking, which for anything genuinely useful is the larger number by a wide margin. Compare total monthly cost, not hosting price, or you will pick the cheap box and be surprised by the invoice that matters.

We do not host OpenClaw. We are not a place to point your existing setup at, and if OpenClaw specifically is what you want — its community, its skills, its way of doing things — one of the three options above is your answer and you should go take it.

What we are is the same wish answered differently. An agent with its own persistent cloud machine, reachable in a chat, that keeps working when you close your laptop. The differences that matter if you are comparing:

The model is included. Claude Code comes installed and signed in, on our subscriptions. No API key, no second bill, no key to rotate — that is the whole reason the two bills above become one.

You do not configure it. There is no compose file, no skills directory, no instance to update. You message a bot and an engineer exists. That is a real loss of control, and for the person who wanted to tune their own setup it is the wrong trade.

One agent is one bot in one chat. Not one assistant with many threads — separate names, separate machines, separate memory of the work.

Own hardware if you have it spare, you want your data at home, and an occasional outage is annoying rather than costly.

VPS if you want control and the best price for it, and administering a server is something you quite like doing.

Managed OpenClaw host if you want that specific agent without the sysadmin work — check the exit, the data and the model bill first.

Something like us if the agent, not the setup, was ever the point.