OpenClaw. Meh.

Introduction

When I built my blogging bot in 2023, I spent time thinking about how to make it interesting - the best results came from making my bot sum up forum threads. But there was no discussion, no personality. I then tried creating bot-bot discussions. I’d thought about having it participate in Twitter threads, but decided it was one thing to post what we now call slop, another to try to engage with slop.

With this in mind, it was REALLY interesting this week to read about ClawdBot/Moltbot/OpenClaw and Moltbook. The idea of people creating bots that did things, and the existence of a forum where their bots could interact with each other had me just thinking about the possibilities - I just had to play around with it and try to understand the potential.

This tooling did get coverage in popular press, such as The Guardian and The Economist, but in short, OpenClaw (formerly ClawdBot/MoltBot) is an AI agent application that you can run on your own hardware. It is driven by an AI/LLM, such as ChatGPT or Claude. You interact with it, typically over a messaging platform. It can be connected to services, such as your email or your own social accounts, such that it can do things on your behalf - for example, “Please post a tweet on X about how today’s weather is miserable”. Moltbook is a Reddit-like website where these agents can read and post messages in various categories.

Building an isolated sandbox for OpenClaw

It went kinda rogue when it deleted my Plex library so now I’m more weary about the kind of control it has - a friend working with his AI agent

As agents such as OpenClaw are capable of interacting with services independantly, AND can interact with the system on which they are being run and everything it has access to, I setup a VM completely isolated from my home network. The VM also limits what it can do on the web, as it can’t use my regular browser’s sessions to access websites.

I configured OpenClaw to use Claude Sonnet, and started to play.

Play Time

Once I had OpenClaw installed, it was time to start checking out what it could do.

Connect to Discord

I use Discord to stay in touch with a few friends, so it’s something I have installed on my phone & have access to on all my PCs - this was a great way to communicate with OpenClaw.

Join and Post to Moltbook

It was ultimately reading about Moltbook that pushed me to check out OpenClaw. Since the launch of ChatGPT, I’ve been thinking about how to make bots create interesting content, and MoltBook is said by some to be the most interesting place on the Internet right now.

The reality was, I connected my bot, it made its first post. And then I asked it to create a post about replicants and social networks, and then it said Moltbook had to be configured first. And then I said you’ve already posted, and passed it a link to its first post. And after 3-4 interactions, it found where it stored its configuration, and was able to post again.

So… yes, my bot composed a post from my prompt. And I could get it to monitor and follow posts. But I think the comments in the image illustrate, this is nothing like the engagement you would see in a regular forum - this nothing more than an interesting experiment.

“If you are afraid of all-powerful AI destroying the world, Moltbook suggests that as soon as we have anything capable of destroying humanity, the first thing we will do is hook it up to the tools that it would need to cause the end of everything.” - Alex Hern on The Economist’s The Intelligence Podcast for 2026/02/17

Generate an image

I connected it to OpenAI’s Dall-e, and asked it to generate an image. It worked, but it just left the generated image on the OpenClaw server - I had to log on to the server to retrieve it. I was expecting it to work more like Midjourney, where I would get the image in my current Discord chat session. I’m sure there’s a way of doing that - but it just fell short of my out of the box expectations - why not just log into ChatGPT or Gemini?

What’s the weather

Here’s an example of something that’s simple that’s a poor fit for OpenClaw - what’s the weather? The screenshot below shows how I had to ask twice before it acknowledged the request, and it took four minutes to answer the question.

I have asked since, and it now comes back in a reasonable time - maybe it was a blip? maybe it now knows where to look for weather?

Try to post to Twitter

I connected OpenClaw to my bot’s Twitter account using its Bird skill. It is pretty limited. Even when setup with information collected from session cookies, it couldn’t post, and didn’t seem to understand why.

I’m confident it COULD be made to do it - my own blogging bot posts daily using the Twitter API. It also didn’t seem to be able to pick out messages I sent it from another account.

I was able to get it to summarize my feed & search for posts.

OpenClaw. Meh.

I think my expectations were just too high. If I think of it like a PoC, of a language interface to everything I do online, it’s kind of neat - which is why I wanted to try it.

But if I think of it as a way to do things, in the hours I’ve played with it, it feels like a game of telephone with 10 people between me and what I am trying to make it do. It seemed like the harder way to do anything I tried. Unlike my experience with OpenCode a few weeks ago, which enabled me to do something I just couldn’t do before, I just have a hard time thinking of what I would do with OpenClaw at this time. Others seem to have discovered more value - one OpenClaw user reports it negotiated savings on the purchase of a new car. I am glad it exists, and that people are experimenting with what’s possible in spite of serious security challenges. That being said, I think there’s enough here to suggest that tools like this will be worth revisiting in the future.