Subject: "OpenClaw"

OpenClaw has been all the buzz lately in and around the AI space. If you aren't aware, OpenClaw is a harness that allows AI models to access 'skills', letting them perform various actions on your computer. The premise of an AI managing your computer sounds like something from a sci-fi novel, something akin to HAL-9000, and that got me interested.

Gemma 4 was also recently released by Google. The western competitor in the self-hosted AI space, It is one of the most technologically advance open-source model made till date, outperforming those that have nearly double or triple the parameter count. This is all possible by using active parameters that gives it a greater grasp on reasoning as well as native support for tool calling.

I won't go over most of the set-up related stuff but reading through the documentations more or less brought it to a usable state. OpenClaw is best if it's tailor-made to your system because everybody has different needs regarding security and automation. I ran OpenClaw inside virtual machine and used LlamaCPP to host the model.

Tooling Around

So the way OpenClaw works, it gives the AI a set of tools. When you first load OpenClaw, it loads a descriptor for each tool and when to use them. For each prompt you enter, the AI model analyses the prompt, checks what tools are available and selects the right one for the job. OpenClaw then receives this tool call and performs it as specified. It then sends a status response back. If other tools are needed, this feedback loop will continue until task is complete.

Let’s say I have a text file. It has a randomly assorted list of fruits. I told OpenClaw to replace all the instances of Oranges to Apples. What happens is Gemma will receive the input, process it, reason and try to find a specific tool. In our example, the model would ideally select the shell tool. It would then select the extra arguments needed by the tool. So it would reply to OpenClaw telling it to use the Shell Tool with sed (Stream Editor) as the command needed for example.

You can also connect OpenClaw to a self-hosted SearXNG instance. You have to configure it to give responses back in JSON format. This allows the AI to access the internet without restrictions. There are other APIs you can use to access search but these have restrictions and rate limits. You would need to configure and run SearXNG in a docker container though.

My Thoughts

As impressive a tool OpenClaw is, it isn’t a be all end all solution to automating systems. Setting it up consumed a good chunk of my day but I did make a lot of mistakes. Some settings were just unnecessarily hidden within config files. For the amount of work setting it up entails, the quality of the responses depend heavily on the model you are running.

Using a capable model would require stronger hardware. Robust hardware requires far more electricity to run. None of this is going to come cheap. For the vast majority of people, keeping OpenClaw running locally around the clock is just not viable. Sure, you can use APIs from cloud-based models but those will also incur the same if not greater cost over time.

I like to think OpenClaw as a good proof-of-concept of what’s to come. Having OpenClaw check devices for heartbeats, do simple maintenance work is actually pretty effective. Unfortunately, I can totally see delusional companies cutting down SysAdmin roles and replacing them with AI. Despite the security flaws, companies would still prefer having a robot that can replace the work of 2 or 3 humans.

The AI bros hyping OpenClaw and peddling AI tech are completely insufferable in my opinion. Most of them are trying to peddle their own AI product to you so of course they have vested interest to push this narrative. OpenClaw is a genuinely cool piece of technology but it's not going to revolutionize everything for everyone like Sam or Dario might say.

For me, I’m happy that the current narrative has shifted away from replacing creativity to instead do mundane work. AI art and videos aren’t going to disappear overnight but seeing projects like Sora fail is the smoking gun the industry needs. The novelty of AI is finally wearing off, becoming just another tool in the box for many.

← Return