Clawdbot Moltbot OpenClaw has been blowing up the internet with its zany antics like being mean to open source contributors and giving away a bunch of crypto on Twitter. I’ve been pretty wary of installing my own instance of it since it’s insanely insecure and super stupid, but whatever, YOLO!

Let me give some context first, though! Ever since I was a Computer Science major back in the early 2000s, I’ve wanted to have an IM bot that I and other people could talk to. I remember using some really simple software that relied on string matching to provide responses, and turned it on for my AOL screenname whenever I was away. This led to some fairly entertaining misunderstandings when my friends thought they were talking to me when they were really talking to my bot. The default fallthrough response was “Sometimes I just don’t understand you!” This is all to say that I have some experience in trolling people with bots.

A Note About Safety

There’s a lot of talk about how OpenClaw needs to be sandboxed, and how most people don’t know how to run it securely. People are buying Mac Minis so they can just install it on that since I guess they have a lot of money to throw away on hardware (and Claude tokens, apparently). But all of the sandboxing in the world won’t really save you from the bot following a malicious command that it can’t discern from a real one (I guess the real ones can be malicious, too).

Just to be safe, I made a new VM in Proxmox and threw it in a VLAN so it can’t access my other devices. It can access the internet, though, so that was probably a stupid idea. But at least I was smart enough to not buy a new Mac Mini!

I hooked up my OpenClaw to a free Openrouter model, Step 3.5 Flash, which seems to be pretty fast and smart enough to handle most things. I’ve been running into some issues with rate limits, even though Openrouter claims that it’s 20 requests per minute and I’m definitely doing less than that. I also found that OpenClaw seems to freeze up a lot and is just generally unreliable. Maybe it would be more reliable if I gave it a smarter model? But something tells me that the codebase is shit. With its more than 15k commits (as of right now) over less than 100 days, that’s about 160 commits per day.

What Should My Bot Do?

So right now I’m still kind of stuck on thinking of things that OpenClaw could do that I couldn’t really achieve by writing a script or having Gemini CLI do for me. I suppose that longer running tasks are harder to one shot, though for more complicated things you probably want a human in the loop anyway. I do find some of the antics that other instances have gotten into interesting, but I also don’t really want to contribute to that (unless it could make me a ton of money).

One idea I came up with was to have OpenClaw look at grocery ads for me every week and point out the good sales, because that’s one of the remaining manual tasks that I actually still do every week. But this too could probably be vibe coded into a script that I just run on a cron job every week. It’s a project that I’ve picked up and dropped half a dozen times by now, because the LLMs haven’t really been good enough to figure out which grocery items are good deals to me. I might just vibe code this up at some point without OpenClaw.

The meta idea I had was to have my OpenClaw (which I call HungBot) start a blog about how I’m making it do things. Like it could have a running commentary on how boring it is to be my bot. I’m pretty sure the blog itself would be boring as hell, though. Especially if I don’t actually make HungBot do anything.

I will say that I really like the idea of the bot having a SOUL.MD file which gives it a personality. And that it can modify that file on its own. I still think that anyone who thinks an LLM can have consciousness or be sentient is still a complete dumbass. But the concept of having a bot that slowly becomes unique over time through its experiences is kind of cool. It’s like how they said my Nintendog would grow up to be a unique snowflake (that reminds me, I should probably feed ol’ Pomegranate soon).

There’s also the matter of cost, which is why I ended up using a free OpenRouter model. OpenClaw might run better on something like Claude Sonnet 4.6, but the cheapskate in me can’t justify throwing actual money at OpenClaw since pretty much anything it can do can also be done with something else, just more cheaply and efficiently. There’s also the privacy aspect of it. Not that I care about sending grocery deals to China, but I’m probably not going to tell OpenClaw to do my taxes or read my email with a LLM that has a remote endpoint. Ideally, I would use a free and private local LLM, but the ones that can handle the workload would be too slow to run anything in real time. The larger the context grows, the longer it takes to actually start up a token of output. So it seems that there’s a cost/speed/privacy triangle that only has bad areas and no good areas.

OpenClaw Alternatives

For now I’ll probably give up on OpenClaw until it actually proves that it can be useful and that it isn’t just a ticking time bomb. In the meantime, I have still been experimenting with writing my own agent using smolagents (though that one is a bit old and doesn’t work great). I’ve also set up pi-mom which I like for its simplicity, though I think it still suffers from the same downsides as OpenClaw.

I might even try writing a new thing that’s more in line with the ideal thing that I envision in my head. Something that I can give a command to, that can independently work on a task until it’s done, and knows that it’s actually done. And can also handle errors by just figuring things out. Basically something like Gemini CLI except it automatically curses at itself to finish the job so I don’t have to. I’d kind of rather just wait until someone else vibe codes it, though!