Starting point: A colleague asked me on Friday morning: “Do you know OpenClaw?” After about ten minutes of internet research, everything was clear again and my weekend was at least partly gone. “Total escalation” on my side.

What is OpenClaw? After Clawd and Moltbot, now OpenClaw. In short, OpenClaw is an open source framework for running your own AI agent with direct access to a local system. The idea is not just to answer questions, but to actively execute tasks. This ranges from simple system commands to more complex workflows. Everything runs locally and stays under your own control, at least in theory.

What happened over the last 48 hours? By coincidence, I still had an old Mac mini on my desk. I had already set it up as another local system for testing small language models. A quick check showed that OpenClaw can be installed on macOS with a single command. The subsequent, more or less complete configuration was done via the terminal. My local Ollama instance, already running on the same machine, could be integrated into OpenClaw without major issues after a short adjustment phase. To communicate with my own assistant, I set up a Telegram channel. At the moment, this seems to be by far the most practical option. OpenClaw provides several commands that can be used directly in the Telegram chat, which makes interaction with the bot quite convenient.

First tests and observations? The first tests using different skills showed one thing very clearly. Since I tested exclusively with Ollama and local models, the chosen model has a direct impact on how well, or how poorly, certain features work. In the end, I settled on gpt-oss:120b. Especially the LLaMA models, which usually deliver good results for me in other use cases, performed surprisingly badly with OpenClaw.

So, is everything great? Not entirely. Some things did not work as well for me as they appear to in various YouTube videos. This is most likely related to the model being used. That said, setting up and running your own AI bot via Telegram works extremely well. From a security perspective, however, OpenClaw is a disaster. The agent has full access to the underlying system. Many bloggers and YouTubers have already pointed this out. For me as a tech enthusiast, nerd, programmer, software developer … the possibilities are still impressive in a positive way.

What comes next? The project is still very young. Even so, it is already exciting to see where OpenClaw is heading and how it will evolve. For me as a geek, this is simply wild. 🤖