Yesterday I wrote about Hetzner’s new AI inference API. While creating the API token for it, a second menu item on experiments.hetzner.com caught my eye, and it is quite something: Hetzner is offering hosted OpenClaw instances there. A fully set-up, always-on AI assistant on Hetzner infrastructure, three sizes to choose from, free for the duration of the experiment.
The remarkable part is not just the offering itself but the silence around it. As of today there is no announcement, no Hetzner blog post, no documentation and no coverage, not even a Reddit thread. The thing is simply there. I spun up an instance and wrote down what I found.
What Hetzner Experiments actually offers
At experiments.hetzner.com, the area where Hetzner currently tries out new things, OpenClaw instances come in three sizes. The names stay within the project’s crustacean theme: Shrimp with 1.5 vCPUs, 2 GB RAM and 8 GB storage, Crab with 3 vCPUs, 4 GB RAM and 16 GB, and Lobster with 6 vCPUs, 8 GB RAM and 32 GB storage. Below them sits the decisive sentence: “This is an experiment, bugs and data loss may occur, free of charge for the duration of the experiment.”
For context, in case OpenClaw passed you by: it is the open-source assistant framework Peter Steinberger started as a weekend project in late 2025, which then took off like few projects ever have, now at over 240,000 GitHub stars, with its own non-profit foundation and a community registry of several thousand skills. The idea: an assistant that runs permanently on your own server, is operated through WhatsApp, Telegram, Slack or Discord, and connects to the real world through skills, from calendars to smart home lights to surveillance cameras. Until now you needed your own server and a good hour of setup. Hetzner turns that into one click.
The quick test
The instance was up quickly, the assistant reports for duty cheerfully (“fresh out of the box, no memories yet”) and responds briskly. The interface strongly resembles the current desktop apps of the big AI providers: chat in the middle, sessions on the left, a context and cost display at the bottom. The settings offer context profiles (from minimal to full), adjustable thinking depth, model configuration and messenger connections, with WhatsApp, Slack, Discord and Telegram ready to pair.
Photo: jonasleupe / Unsplash
The skill catalogue reads like a tour through the community: evaluating camera streams via ONVIF and RTSP, controlling Philips Hue lights, Google Workspace integration, PDF editing by voice command, email triage workflows, spectrograms from audio files, even Sonos and Spotify. In equal fairness: a considerable part of it is currently locked in the hosted instance or fails on missing prerequisites, and the assistant’s web access did not work at all in my test. The experiment is visibly at a very early stage.
That led straight to the best stumble of the test. Asked what the instance actually costs, the assistant, lacking web search, dug through its own documentation, worked through 38 tool calls and then confidently informed me that I had rented a paid cloud server with a monthly fee. That is wrong, but it sounds entirely convincing. You could hardly wish for a better demonstration of why AI answers without reliable sources deserve scepticism.
The confusing cost question, resolved
Two things about this offering systematically create confusion, and both deserve a clear resolution.
First, the button. Creating an instance or an API token means clicking “Create and Buy Now” without any price shown anywhere. This is recognisably the standard order flow of the Hetzner console, simply not adapted for the experiments. Under EU consumer rules the matter is clear: an online payment obligation only arises when the price is clearly displayed before the click and the button unambiguously states the obligation to pay. A “Buy Now” without any price on an explicitly free experiment does not create a purchase contract, and should Hetzner later turn the offering into a product, that requires a new, priced offer which you actively accept.
Second, the cost display in the interface. OpenClaw shows estimated costs at the bottom of the chat, and in my test a single research question produced an estimate of around 60 cents. That number does not come from Hetzner but from OpenClaw itself: the software multiplies consumed tokens by the API list prices of the configured model, regardless of whether anything is actually billed. And the token numbers explode quickly, because every tool call counts the entire context as input again, which is how a “hello” followed by one documentation lookup racks up six-figure token counts. It is a display for perspective, not an invoice.
A third curiosity on the side: my Lobster instance reported 12 CPU cores and over 60 GB of RAM in the system overview. The reason is mundane, the container reads the values of the shared host machine instead of its own limits. The booked 6 vCPUs and 8 GB are what you can rely on.
Who this experiment is for
The offering is interesting above all for people who always wanted to try OpenClaw and got stuck on the server or the setup. Exactly that hurdle disappears, free of charge on top. If you want to know what an always-on assistant with messenger integration feels like, this is the easiest entry there has ever been.
Photo: rev3n / Unsplash
Equally clear is what you should not do with an experiment instance, namely entrust it with anything important or confidential. Hetzner itself warns about bugs and data loss, and OpenClaw as a project has a security history worth knowing: in early 2026, 341 malicious skills were discovered in the official community catalogue, followed by several serious vulnerabilities in spring. The project is visibly working on it, but real mail accounts, customer data or server credentials have no business in a sandbox like this. For getting to know the concept with non-critical tasks, which is what a free experiment is for, there is nothing wrong with it.
Strategically, all of this fits the pattern I described yesterday with the inference offering. Hetzner is currently testing in several places at once how much demand exists for AI infrastructure out of Germany, first the model API, now the hosted assistant on top. Together they would eventually add up to a complete stack on EU infrastructure, and that would be no toy anymore.
If you are wondering what an always-on assistant could sensibly take over in your business, say pre-sorting incoming enquiries while nobody is in the office, feel free to get in touch. I build automations like that without OpenClaw too, tailored to the specific case and without your data ending up in an experiment.
FAQ
What does the hosted OpenClaw instance at Hetzner cost?+
Nothing during the experiment phase, and it says so explicitly: free of charge for the duration of the experiment. The button is labelled Create and Buy Now and the OpenClaw interface shows estimated costs in cents, but both are misleading. The button is the standard Hetzner console flow, and the cost display is the software's internal projection of tokens times list price, not an invoice. A payment obligation can only arise later with a clearly stated price and your explicit consent.
What is OpenClaw anyway?+
An open-source framework for an always-on AI assistant, started in late 2025 by Peter Steinberger and gone viral in early 2026, now with over 240,000 GitHub stars and its own foundation. The assistant runs on a server or a Mac, is operated through messengers like WhatsApp, Telegram, Slack or Discord, and gains capabilities through a community catalogue of several thousand skills, from camera monitoring to smart home control to email triage.
Should I trust the instance with real credentials or documents?+
No, for two reasons. First, Hetzner itself warns about bugs and data loss, and the experiment can end at any time. Second, OpenClaw as a project has a security history that deserves respect: in early 2026, 341 malicious skills were discovered in the official community catalogue, followed by several serious vulnerabilities. The instance is meant for trying things out with non-critical tasks. For mail accounts, customer data or server access it is the wrong place.
Why does my instance show more CPU and RAM than I booked?+
The instance runs as a container on a shared host, and the system display in OpenClaw reads the host machine's values instead of the container limits. If you pick the Lobster size and see 12 cores and over 60 GB of RAM, you have not received a bigger machine. What is guaranteed are the booked 6 vCPUs and 8 GB, the rest belongs to the neighbours on the same host.
Want to know more?
In a free intro call we discuss how you can use these topics for your company. Not a sales pitch, but an honest assessment.
Book a free intro call
