Where to host OpenClaw: self host vs managed, every option compared

Once you have decided to run OpenClaw, the next question is where it lives, and it matters more than it sounds. A personal agent is only useful if it is awake when you need it, so it has to run on something that stays on around the clock.

That rules out the obvious candidate, your laptop, which sleeps the moment you close it. What is left are a few real options, and they trade off cost, control, and how much of your time the upkeep eats. Here is each one with current prices.

What hosting OpenClaw requires

The requirements are modest as long as the model runs in the cloud. OpenClaw orchestrates the agent and calls a hosted model for the thinking, so the machine itself only needs to keep a small Node service alive.

You install it with npm or Docker, supply a model API key, connect a channel, and then keep the box patched and the gateway running. The one case that changes the math is running a local model through Ollama, which wants far more memory and a real GPU; that is a different class of machine and most people do not start there.

The baseline for a cloud model setup, in short:

A VPS

Renting a small cloud server is what most self hosters reach for, because it is cheap, always on, and someone else owns the hardware.

The prices have spread out a lot. Hetzner is the value pick: a CPX22 with 2 vCPU and 4 GB of RAM runs €7.99 a month in Germany and Finland after the April 2026 price adjustment, up from €5.99, with a 1 vCPU box cheaper still. DigitalOcean charges around $24 a month for the same 2 vCPU and 4 GB droplet, which buys you a more polished console and more regions but is roughly three times the price for the same compute, as side by side tests keep finding.

The catch with any VPS is that you are now a small systems administrator. You manage the operating system, security updates, the OpenClaw install, and the gateway when it falls over.

A rented box also comes with a public IP, so the gateway is reachable from the internet the moment it boots, and part of the job is closing that down: turn on the gateway's tokens and allowlists, put a firewall in front of it, and follow the security guide instead of leaving the port open to anyone who scans it.

Budget hosts generate a steady stream of forum threads from people whose WhatsApp connection or gateway broke on a shared box. The rent is cheap; your evening debugging it is not.

A machine at home

If you would rather buy once than rent forever, an always on computer at home works well. A Mac mini is a popular pick because it idles at three to four watts, which works out to roughly $15 to $25 a year in electricity, and it runs silently for years. The hardware is a one time cost and the running cost is trivial.

The Raspberry Pi used to be the obvious cheap option, and in 2026 that is less true than it was. A memory price spike, driven by the same AI infrastructure buildout that makes agents interesting in the first place, pushed Pi 5 prices up sharply: the 8 GB board now sits around $175 and the 16 GB around $305, where the 16 GB launched at $120.

For OpenClaw on a cloud model you do not need much memory, so a 2 GB or 4 GB Pi 5 at $65 to $110 still works and sips power. But a Pi is no longer dramatically cheaper than a used mini PC, and a second hand Intel or AMD mini box with 8 to 16 GB of RAM often lands in the same range with more headroom, at the cost of a few more watts.

What you take on with any home machine is everything a homelab implies. Your home internet and power become the agent's uptime, so an outage or a reboot takes the agent down with it, and you are still the one applying updates and restarting the gateway. For a single personal agent that is often fine, especially if you already keep a machine running at home.

Reaching it remotely is worth setting up too, with something like Tailscale, so you are not stuck when it needs a restart and you are out of the house. Keep that access on the private network rather than forwarding the gateway's port to the world.

The gateway is the door to your agent and every tool it can run, so reaching it over a private mesh means only your own devices can connect and you never put a login in front of the whole internet.

Docker

Docker is not a place to host so much as a way to host. Running OpenClaw in a container tidies up the Node and dependency wrangling and makes the install reproducible, and it works on a VPS, a home server, or your own desktop.

It removes a category of "missing git binary" and "wrong Node version" install pain, which is real. It does not remove the need for an always on host underneath it or the job of keeping that host healthy. Think of it as making whichever machine you picked easier to manage, not as a fourth machine.

A managed host

The last option is to not host it at all. Operator is managed OpenClaw: it runs the framework on its own infrastructure with a model, web search, your channels, and connectors already set up, so you sign in and the agent is running.

You pay a flat subscription on Operator pricing (Basic $20, Pro $50, Max $175) that folds in hosting and included AI usage tiers, and there is no VPS to patch, no gateway to restart, and no Node version to match.

The trade is straightforward. You do not own the box it runs on, and in return you do not run the box. For people who reached for a cheap VPS to get an agent and found themselves debugging npm instead, that is the whole appeal.

Model cost

Whichever box you pick is rarely the part of the bill that matters most. A VPS runs from a few dollars to about twenty a month and a home machine is a one time purchase, but the model behind the agent is the recurring cost that scales with how hard you work it.

When you self host, you bring your own model key and pay the provider per token, so a chatty agent doing real jobs can spend more on inference in a month than the server costs to rent. Running a local model on your own hardware removes that token bill, in exchange for a much larger machine and models weaker than the frontier ones.

A managed plan usually folds the model into the subscription, which is part of why a single flat price can come out ahead of a cheap VPS once you add the API spend on top. Compare the model cost alongside the rent on the box, not on its own.

How to choose

Here are the options side by side before the recommendation:

The decision comes down to what you would rather spend. If you enjoy running servers and want the lowest possible bill for raw compute, a Hetzner box plus your own model key is hard to beat at well under ten dollars a month, and a machine you already own at home is cheaper still to run.

If owning the hardware and the tokens is the point, self host and treat the occasional install error or gateway restart as part of the deal. If your time is the scarce resource and you want the agent to just be there, a managed host removes the entire operations layer for a single subscription.

Plenty of people start self hosted, hit the upkeep wall, and move to managed; others go the other way once they want full control. You can try the managed path free before deciding, or read what OpenClaw is first if you are still mapping the pieces.