For internal use cases, we are constantly developing and testing new and exciting approaches that we’d love to share with you.
To give you access to our prototypes, ideas and experiments, we’ve launched the Experiments platform.
We invite all Hetzner customers to try out our solutions, which we host on our own hardware in our data centres. No guarantees, but no cost either.
We offer an inference API that provides access to an Open-Weight LLM via an OpenAI-compatible REST API. You can use the API with any OpenAI-compatible development environment and easily integrate it into your own projects.
You can find a setup guide for OpenCode in the Community Tutorial. We pre-configured the LLM, Open-Weight. The quality of the responses is probably not as good as that of the major providers, but it’s often sufficient for initial experiments.
Building on the Inference API, our second experiment is an OpenClaw platform that we host entirely, including our own pre-configured models. You can configure your own OpenClaw exactly as you wish and also connect other inference providers. Try out the OpenClaw AI assistant as an alternative to Claude or ChatGPT and retain control over your data.
A word on OpenClaw in particular and AI in general: be aware of the risks when you provide your data to OpenClaw. We have no influence over the OpenClaw code and accept no liability for OpenClaw’s behaviour. In the event of high demand, our experiments may no longer be available or may not perform as expected.
Because we do not create backups, you should back up your configurations separately and avoid setting up production environments on the Inference API.
We hope the Experiment Platform will provide us with both concrete feedback and opportunities to shape the project, as well as various types of metrics: How well does our approach scale?
Is there a need for and interest in Software-as-a-Service offerings operated by us? Which (additional) features are particularly important to you? You can also deliberately test the limits in this experiment. This provides us with valuable data, showing, for example, what load our setup can withstand.
Please be sure to share any feedback, questions or experiences with us via a support request from within Experiments, or feel free to discuss them on the Hetzner Forum.
Completed experiments
Here you will find a simple overview of experiments we have already run and finished. These experiments serve as a reference and show which approaches we tested and what results we achieved.
Experiment 1: OpenClaw
We tested the OpenClaw AI assistant, hosted entirely by Hetzner.
Objectives
- We wanted to give users an OpenClaw that was ready to use and secure, with no extra setup.
- At the same time, we aimed to gain first-hand experience with self-hosted inference. OpenClaw provided a controlled environment for this, as inference was only accessible through it.
- We also wanted to test the Experiment Platform concept ourselves to collect early feedback from users.
Results
- The code quality of OpenClaw is too low and varies too much.
- This would require too much support and maintenance.
- Users prefer to connect to inference directly with other clients or harnesses, without OpenClaw in the middle.
- This led to the next experiment: a publicly accessible inference service.