Recently, I created my own version of OpenClaw in KNIME. A friend suggested that others might find it interesting too—if there’s enough interest in Like Bottom, I’ll make a video showing how I did it.
Cheers,
AG.
4 Likes
From starting to understand how agents / tools work to replicating OpenClaw: That sounds like a pretty steep learning curve ;-). I’d definitely be keen to see how you have done that!
Hahaha, it’ll certainly be a steep curve—but I have to disappoint you. I’m not just starting to understand how agents work; I’ve been working with them for over 1.5 years, directly in Python using LangChain and n8n. I’m actually a firm believer in agentic systems as the path to AGI—I told Yann LeCun exactly that when he was in Tübingen about a year ago, even though he wasn’t convinced. It’s not about LLMs alone—it’s about agentic collaboration, which produces emergent properties. I even think OpenClaw is the first small step toward “real AGI.”
I’m new to KNIME Pro and the server, but that didn’t stop me from building the full infrastructure to run this workflow while keeping the KNIME Analytic Platform open on my computer. I’d love to show you—just need to find the time to record a video, if people really want it. It’s a bit of work, but totally worth it.
2 Likes
Just to make sure you don’t get me wrong - I did not mean that you are a general “agent beginner” (how could I possibly know), but was more refering to an earlier post you made in relation to how to build tools within the KNIME AI Agent framework :-).
As I said I’m keen to find out. I have a fair bit of catch up to do on KNIME AI Agent functionality - e.g. the examples on how A2A was implemented using KNIME Hub Service deployments etc…
Luckily one of my private projects “just” wrapped up so hopefully more time for things like that soon!
It is a fair call. I am not sure that using KNIME tools will be even possible to do a multi-tool call with the agent in a single inference event. Perhaps you can tell me. We have our own MCP server and the call tools are routed to those. But it will be a great update to use the KNIME tools for analyzing the tabular data using KNIME tools strategy.
After i figure out how to properly deploy in the Server complex workflows, i could upgrade “Jarvis” to do this as well.
Lets see is the time let me do this soon.
1 Like
that’s an interesting idea. What would be really interesting to see how you control the agents in regards to token consumption and cost, because I think this is one of the keys making this really scalable.
OpenClaw per se is interesting but if you run with a standard Opus 4.5/4.6 model, this becomes extremely expensive very very fast (I burned through $20 in under 18 hours and also replaced some functionalities with local Ollama and Sonnet models).
I wonder how you approached the heartbeat feature, is this is a “simple” (for the sake of a better term) schedule?
What also would be very interesting to see how you “free” the agents to have full system access (if that’s what you’re building).
So in a nutshell this is me saying “yes, I want to see this” 
2 Likes
I’m very happy to hear this feedback because these are exactly the challenges I’d expect from people starting to work with second-brain level AI systems.
We’ve been using a very similar system inside our company for quite some time now, so I completely understand the pain points you’re describing – especially the cost issue with closed-source models.
The good news is: all three of your questions (token/cost control, heartbeat implementation, and system access) are exactly what we’ve been solving and refining along the years. We’re really happy with how it works.
I’m not going to spoil the solutions here in the comments, but if there’s enough interest (you know what button to press 
), I’ll make sure the video covers all of these practical concerns in detail.
Thanks for the questions, that will help me for the video.
Cheers,
AG.
1 Like
It’s amusing to see how many people viewed the post and how many clicked “Like.” I’m curious whether any of those 103 interactions are bots—perhaps surveying content related to OpenClaw—or if they’re genuine users. To me, at this point, it is clear. 
1 Like