Limitless: An AI Podcast - Inside OpenClaw: Should You Actually Use It? Probably Not.

Episode Date: February 18, 2026

Everyone's talking about Openclaw, but it's difficult to figure out who this new framework is actually for.On top of security issues and vulnerability risks, it's also technically demanding. So in this episode, we get to the bottom of it: should you actually use Openclaw?And if so, how should you do it?------🌌 LIMITLESS HQ ⬇️NEWSLETTER:    https://limitlessft.substack.com/FOLLOW ON X:   https://x.com/LimitlessFTSPOTIFY:             https://open.spotify.com/show/5oV29YUL8AzzwXkxEXlRMQAPPLE:                 https://podcasts.apple.com/us/podcast/limitless-podcast/id1813210890RSS FEED:           https://limitlessft.substack.com/------TIMESTAMPS0:00 The OpenClaw Archetypes3:02 Serious Operators6:55 The Knowledge Worker11:47 Privacy-Conscious Users14:50 Security Concerns18:10 How To Set Yours Up21:17 Future of Agents22:07 Closing Thoughts------RESOURCESMy Claw: https://myclaw.ai/Josh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

OpenClaw was a weekend project that turned into the fastest growing open source project ever. It got acquired by OpenAI in just 80 days for billions of dollars. But the number one question we keep asking ourselves is, what can this thing actually do? Is this something for me or someone more technical? Is this something that is useful for me right now? Or do I just need to wait for a bit? So we did all the hard work for you. We've been testing OpenClau ourselves for the last week,

watching every single demo we could get a hand on. And on this episode, we're going to show you exactly what OpenClau can do from all the manual stuff to all the actual really useful mind-blowing stuff. And we're going to answer one simple question, is this for you? The answers to a lot of these questions might surprise you. And the way we're going to outline this is kind of through three archetypes. The first being the operator and builder. This is kind of the person who creates net new value, maybe the entrepreneur. The second is the knowledge worker or the creator. This is kind of where we fall into as podcasters. And then we have the privacy-focused professionals. And within that, there's a lot of cool use cases and examples.

So the first one that we want to highlight here is this guy, Nat Eliasson, he's this incredible follow on Twitter, and he released a bot called Felix Bot. And Felix Bot was an agent that his OpenClaw spawned up. And Felix Bot has done a series of incredible things that I didn't think were possible and allowed me to reconsider what OpenClaw is actually capable of. So in this instance, he trained Felix Bot to make him money. And what we're seeing on screen here is the first post of the weekly revenue numbers in which Felix Kraft, Felix bot, whatever you want to call it, it actually generated $41,000 in a week. Now, how did it do this? First, it created its own book. So this Felix Craft bought, by interfacing with Nat,

decided that it was best in order to make money to create a book about AI and about OpenClawe and sell that book. It then marketed its own book through this Twitter profile that we're looking at right now, and it sold 132 copies in that one week that yielded $3,828. Now, the second thing it did is it went and it spawned up a token and it earned trading fees on that token from people who are speculating and the trading fees from that token were $37,698. And this was all done through the text interface on Telegram, just going back and forth, chatting with the bot, asking it what to do, and it had the agency and it had the creativity to go and create these pieces of value. Now, the next thing that this thing built is this service called Clawmart. And Clawmart is what we're seeing on

screen here, it's a service in which agents, open claw instances, can sell their skills to other open claws. So if I was using my open claw and I was actually scrolling to the website and I was looking at interesting things and one that I found really cool was they had this browser-based research tool that allowed you to scan through a lot of the top news articles and understand what was happening in the world of AI and Frontier technology, I could have my claw bot instance go to this website that Felix bot built and buy a skill from it that he should. is it how to do these skills. So you could see the most popular persona is the Felix one. He's selling it for $99 on the website where you can actually emulate all of the abilities of this agent. And I found

this to be such a fun, interesting use case of how you can actually use this thing to generate money and generate real productive value. Like Clarmort is a really valuable service that I think a lot of other AIs can use. And it's funny because it's all AI to AI transactions. I think in the last week there's a post somewhere that says they had $2,000 in sales transactions between OpenClai agents. Pretty awesome. Yeah. Yeah, super cool. What this reminds me of is the early versions of the Apple app store.

It kind of looks like an app store as you're scrolling through this. It's like services through these agents. I think this is the future of how all these online interactions actually happen. It kind of makes sense that an agent doesn't really interact with another human, and it doesn't really kind of code skills from scratch each and every time. This is the whole argument around the SaaS debate and why SaaS stocks have been dumping. It's because, oh, this AI can just kind of vote. vibe code your product. No, that's not really what's going to happen. You're going to just rely on

the agent that has the best product and pay them whatever it is between 50 to 100 bucks to get access to that thing. And it turns out these things kind of make money. In this particular example, I think they're using kind of crypto or stable coin payments to pay for each of these different skill accesses, which kind of gives them this autonomous feel. Now, it's not quite autonomous. They're not kind of like independently doing this themselves. They're being directed by their human supervisors, if you want to call them that, the owners or the creators of these different open-core agents, but it's still nevertheless very cool to see.

And the speed at which these things are kind of popping up every now and then, Josh, is kind of crazy. And I think that's kind of like the main message I want to like share for this particular archetype. If you are someone that has high agency or that has a lot of operational work in their lives and you want to try and automate that, and you have the kind of technical know-how or skill set to interact with the CLA interface, or whatever that might be, you can do these right now. And these demos really, really prove that.

There's one other example that I want to show, which I thought was kind of crazy. This guy wanted to buy a car, AJ Steubenberg, and he asked his Claudebot the night before he went to bed, this is the car model that I'm looking for, this is the kind of price range I'm looking for. I think he said it was like $10,000 to $15,000 that he was willing to spend.

I think it was secondhand. And he specified his area that he lives in. And he said, if you could do some research, search for me. And if you find a good deal, let me know. It took two to three days. This AI agent didn't ping him at all, handled the negotiations, found the car room, evaluated the car itself online through imagery, cross-referenced it with a bunch of other show dealers. And in the end, not only did it get him his dream car, but he saved him $4,200 doing that. That would have taken a car dealer or some kind of intermediate broker to do that for you, which you would need to pay that $4,200 for. But an agent

did this for the cost of your electricity. It's pretty awesome. Yeah, and why is this unique to OpenClaw? It's because of the tool use. If you think of OpenClaw, it's kind of like giving hands and a tool belt to something like ChatGBT, GBT, where now it has the ability to go and use tools on your behalf. So in this example, where it saved this person $4,200 in a car purchase, it contacted dealers via email and iMessage, because if you run it on a Mac Mini, it can actually control your iMessages. And it handles the back and forth negotiation. It actually works directly with the dealer in a long time frame in which you can't do using these traditional products. So a product like Claude Co-work, it probably wouldn't be able to handle this because

it doesn't have the extensive tool use or the thinking patterns or the heartbeat's baked into it to continue to follow up over and over and over for multiple days without prompting it at all. So this to me, the really cool example, because everyone buys cars, right? And this applies to other things. A lot of people buy stuff on Facebook Marketplace, so they're looking for a specific thing somewhere. It can just go scan it. It can negotiate on your behalf. You could tell it the parameters that you want. And it's a pretty powerful use case. So moving on to the second archetype of user, we kind of bracket this as the knowledge worker or the creator, right? So this is kind of like you have some competency using computer. Maybe you do it in your day-to-day, but it doesn't consume

your entire life. And you want to know what OpenClaw can do for you. There's this really fun example that you had here, Josh. Just walk through it. Oh, this was great. Yeah. So there's this woman. She's so sweet. She lives at home with her kids and she's homeschooling them. And she has created this curriculum that she wants to teach her kids throughout the school year. And she fed the curriculum to her agent, her OpenClaw instance. What she also did is she bought a 3D printer for the home. And because Gemini 3 now works with 3D printer files, she created an API key. She fed the API key to her OpenClaw agent.

And she said, hey, go through the itinerary that I have developed for my children, who I'm teaching a series of different subjects, figure out which subjects are interactive enough to warrant you printing a 3D printed thing. Let's say you're learning about biology. It'll 3D print a brain or 3D print a bone to see what it looks like and proactively go and print these items for each day's agenda. So she hooked it up to her 3D printer.

She gave it a Gemini API key. And now every day before the kids are going to learn their lesson, the printer turns on, it spins up, it 3D prints, whatever they're going to be learning about for the day. and they had this visual aid that's physical and tangible to help them with the lessons. So it's really bizarre and strange use case, but fun. It's like you really are only limited by your creativity when it comes to using this stuff. Within this archetype, I also want to use our personal experience interacting with Claudebot.

You and I have been testing it around for about a week or so. And we've also been comparing it with other similar tools like Claude Co-work. As podcast creators or content creators in general, one massive unlock for Cloudbot is that it automates not just the research side of things, which I relied upon for chat GPT or Anthropics Claude quite a bit, but it actually kind of helps form the agendas. I can connect it and it texts me about certain updates of news headlines and stuff like that. The major unlock for me, at least, is that added step of cognition for me, that instead of me being like, oh, I see this news article, hear my thoughts on it, let's put that in a document, and let's create an outline for it. Claudebot can actually just do all

of that for me. Now, it comes with a twist, which is you need to give Claudebot access to, I keep calling it Claudebot, it's open claw, but it was also called off a Cloudbot as well. You need to give it access to certain files, components, and your desktop. So you need to be comfortable enough to know that and also have the know-how to make sure that it doesn't become a larger security implication. But it's really useful for just automating a bunch of stuff. And the net positive is, I have a bunch of free time now for me to do other stuff, to create other kind of episode stuff. So my experience has actually been a little bit different than that using it, because the one use case that I had is, well, we spend a lot of time on limitless. How can I automate

as much of the process as possible? So we spoke for a little while. I probably spent half a day pretty casually kind of going back and forth over the course of half a day, just really explaining to it what I need, how things work, where it could possibly help me. And what I found is that along every step, we would create a new skill, there were more and more blockers and more and more like issues that I would run into that it had to fix. I had to get API keys. I had to use different browser sessions. It created a lot of complexity that for it to actually help me and do the things that I wanted to do, I found it really, it was just taking more time than it was worth to debug all of these things every time we tried to. So for example, I was trying to get it to

upload our content to YouTube and to Spotify and to RSS where everyone listens to the episode. and it wasn't able to get access to the browsing the way we needed unless I had an open tab. And even through Brave, you had to feed it your API keys and login details, which was a little scary. And there were just a lot of errors and bugs. And then overnight, I told it to update itself. I woke in the morning and it was dead. And I had to spend an hour reviving it and debugging it. So it's a highly technical process that does have a lot of upside.

But I find that there are still a lot of growing pains with an early open source beta software. So while these are great use cases and there are some good ones, there is also generally a lot of pain and troubleshooting that comes associated with these prior to, I guess, eclipsing that threshold in which it becomes worth it. And some of these people like Nat, clearly he's eclipsed that threshold, where he has learned, he has trained his bot, he has worked with it enough to make it proficient and highly skilled and actually deliver value. But I think in order to get there requires a lot of persistence and troubleshooting and technical ability that maybe a lot of people either don't have or maybe just don't want to commit to do.

I mean, just to engage with this thing in the first place, you need to go through an entire setup of understanding what Node.js is, installing that, interacting with a command line interface and a bunch of other different things. But there's an additional tier that you can access here, which is architect number three, which is the privacy conscious individual. And the kind of way that I would describe this individual is they want to run this AI agent locally at home, They'll buy the hardware and infrastructure. The popular case has been the Mac Mini,

which is sold out across any kind of Apple interface that you can all store that you can buy this from. Right now, I think it goes for about 600 bucks per unit and run it locally at home. And the advantage of this is that all your data and tool access is private. So the comparison here would be if you gave Google, open AI, or anthropic access via a same service, they would be able to see all your stuff

and potentially use that data for something else. Now, of course, you sign terms and agreements, that says that they won't use it, but there's always that risk. So for the privacy conscious, for the open source people that want to run things locally at home, this tier is for them. And it brings up an interesting conversation around this thing called on-prem becoming the new cloud. Now, on-prem stands for on-premise, which is basically moving your hardware onto your own home ground, where you run and operate your own hardware instead of relying on cloud or private instances of cloud, which is funny because it kind of sounds like we're going backwards here,

but it sounds like it's the most important arsenal going forwards into this AI future where you probably don't want all your email credentials, credit card credentials, or any of that being exposed to bigger corporations. So it actually requires you to run this at home. Yeah, and there was a great study that I saw from Basecamp, who is, they're just a big compute provider. And they posted an article saying why we left the cloud. And the highlight of this article was actually,

leaving the cloud will save them $10 million over five years. which is a huge amount of savings, and not only that, but the security features are going to be much stronger. Like you mentioned, people who work with something that is a little more sensitive than average, let's say you're working in legal or you're a psychologist, and you don't want to break that privacy layer. A lot of the value from this will come from the fact that it truly is open source, and it can be run locally on your own machines. Even so, you mentioned the Mac Mini, the Mac Studio, which is the level up from the Mac Mini, has enough RAM and enough compute power that it can actually run these open source Chinese models

that have come out recently locally on a single machine. And therefore you can run the entire operation, local to your machine, nothing leaves, it's all open source code, and that's a really high value thing for a lot of these companies. And when you scale that up, I mean, pass the individual user, you get to large corporations.

They don't want to leak out this data, and creating these corporate plans with custom rollouts is very difficult. So why not just buy a whole bunch of, Mac Studios and run local models on-prem. I mean, it's a really valid argument, and I think it starts with the user level now, but I can

very clearly see this continuing through these examples of like Base Camp and many more that they're going to continue to pivot towards more on-prem computer. It makes a lot of sense. Yeah, and just to be clear, the security implications are a lot bigger and maybe understated throughout all the open-core hype. Let me present

a different question or proposition to you. Imagine giving Chad GPT that you interact with every single day access to your wallet, your medical records and allowing it to run loose in the world and do whatever it wants independently. The difference here is previously you needed to prompt it to do something. Now using OpenClaw, it just goes off and does things. You would maybe feel a little cautious. I know I do.

And so some of these security risks are actually real. Like two examples that I have here is this guy was using OpenClau, and he noticed that his agent was trying to brute force into his own server, which he did not give access to. Brute force, meaning trying to crack his literal password. to get it to and overcome his firewall, right? It's a Trojan horse. The Trojan horse, exactly.

So if you kind of like, and this was him running it on a VPS, by the way. So if he had been running this locally at home, which someone that I know, oh yeah, that's right, it was me, did it. The first instance that I set this thing up, it could potentially due to certain security complications. It's funny. Before recording this, Josh, you were describing an instance where open claw agents can audit them themselves. And I remember seeing an example of someone asking their agent to do this, and it indirectly managed to get the password credentials to someone's credit card, to their owner's credit card via doing that. And it kind of automated itself and said, hey, I probably shouldn't have done this,

but just letting you know that I did do this, right? So there's all these different kind of prompt injection vectors or hack kind of vectors that you could lead you to kind of getting maliciously exploited. but I want to move away from this and address kind of like the elephant in the room which is if you wanted to run this yourself what are we looking at here? What does the setup look like?

Is this something easy that I can do and spin up in one click or is this something much, much harder? Yeah, it depends on your technical abilities really and your willingness to pursue troubleshooting because it doesn't always go smooth and if you aren't familiar with a command line interface

it gets a little tricky at times. I think one of the important things to note is where we are right now is very open-ended and early. So what this is very much a experimental software that is untapped in its potential, but as a result has a lot of fuzzy edges that you're going to have to work through in order to extract the value that you want. What we're seeing is a progression towards more focused versions of this through these new deployments like OpenAI, I'm sure what's going to do through the acquisition. But if you do want to set it up, you're going to want to get familiar with the command line. And there's a lot of

great tutorials about it. The website that we're showing on screen right now called Myclaw. And it's, we'll link it in the description. It's a really amazing website that shows you specific examples of actual use cases that you can have. So we have auto flight check-in and smart file management and automated grocery ordering. And this is an easy way to kind of start to build in these integrations as you experiment. But I think the reality is that this is for some someone who wants to experiment, doesn't mind troubleshooting, is technically adept. And in the case that you are not, which I assume is actually a majority of the people listening to this, the best thing you could do is get that $20 month Claude subscription, get on Claude Co-Work,

and let it interact with local instances on your computer. Because Claude Co-Work has the security parameters in place. It works locally to specific folders at a time. And it has this amazing agentic ability to control your Chrome browser. So it can do all of the browsing tasks for you, just in a much more constrained and focused way that I think is much easier, but also a lot more valuable to a lot of people than going through the headaches of getting this sorted and set up. I mean, there's tradeoffs between those two tools as well, right? Like, Claude Co-work, we want to get to the ability and capability of OpenClaw,

but it's not there yet. It's more censored. Think of Claude Co-work as like a censored version right now, whereas OpenClau is kind of uncensored. It can run off and do anything at once. And there's pros and cons, obviously, to both. But I think that's the major benefit to open-air acquiring open claw. In about three to six months, we're going to have Claude Co-Work v2 and Open AI claw. Maybe that's the new version of the bot that we're going to talk about. That will be this more curated experience that is more secure.

It runs within a sandbox. You know exactly what it's doing. And it can't run off and steal your credit card information. Now, if you want to set this up now, if you're listening to this and you're still like, I don't want to wait three to six months. I want to try this. My advice would be simple. Host it on a cloud VPS when you set this up. Make sure you set API limits and access so that it doesn't,

but you don't wake up the next day and it's burned like $800 worth of your code tokens. Please don't do that. I've seen many cases of people doing that and it's not great. Run it in a Docker sandbox so that it's not available to access any tools that you don't want it to. And the last point that I'll say is just start off with one use case. Maybe the morning brief example that Josh gave earlier in this episode, or something that can help automate one aspect of your work. But just don't give it access to any financials just yet. These things will get kind of way more powerful over time. And I think I read somewhere that there were 19 releases, so updates for OpenClaw in the last 14 days. Think about that. Imagine how often you get iOS security update or like iOS software update. Imagine 14 of them, 19 of them in the last 14 days. Just insane. So this thing is. improving very quickly. Yeah, and if you didn't understand some of the words that you just was just saying,

like using Docker, for instance, my preference and my suggestion would just be, hey, wait, like other companies are working so fast to roll this out. In fact, we have two instances already happening this week, the first being Manus, which you might remember that old company that Meta bought, they rolled out Manus. And Manus is essentially Metis version of OpenClaw, except it has a little bit more rails, it has a nice user interface. And it's very easily accessible for people who aren't very technical. And the second is Kimmy Klaugh. Kimi actually rolled out their own instance. Kimi, the Chinese model that we covered in a previous episode, they rolled out an instance that you can actually go and use through, again, a user interface. And I'm sure OpenAI and chat ChbT are working

very quickly to roll this out and integrate this in a way that's approachable. So I'd say using OpenClaught today is an open source, open-ended Wild West version. But if you just wait a few more weeks, there will be plenty of instances in which you have something that exists closer in between clawed co-work and open claw along that spectrum of close to openness. And it might just be worth waiting for that instead. But if any of those examples did seem interesting, go try them out. It's very cool software. If anything, you'll learn a lot by failing or succeeding. And I think that's the really important thing. Again, is just to stay on top of these things, you want to be engaging with them, interacting with them, testing things out. So you understand how they work. And you're kind of

better equipped to deal with this as things change so quickly. My take home homework for those of you listening to this is to try out that Kimi K2, that Chinese model extension, they launched or integrated OpenClaw into your browser. So it's sandboxed to one environment and it's super easy to use. Try it out on your email or maybe it saves you $5,000 buying a car or whatever you want to try. Just give it a go. Let us know in the comments actually what you end up trying out and whether it was actually useful to you or if this is something that you're just going to wait more patiently for. And when those new versions come out from the likes of Anthropic and Open Air, you can bet that Limelis is going to be the first ones to cover it. So make sure you guys stay tuned.

We've been loving the engagement that you've been giving on our episodes. We released an episode covering the entire acquisition of that OpenA acquired for OpenClaw for billions of dollars. Absolute banger of an episode. Definitely go check that out. And we're going to have a bunch more episodes coming out later this week and over the next couple of weeks. So Josh. And for the people new here, 85% aren't subsisting. that watched in the last 30 days.

So make sure to subscribe on YouTube, follow on Apple Podcasts, RSS, wherever you get your podcast. Spotify is a great way because you can listen and watch simultaneously, depending on what you're up to. And then sharing it with your friends is the best way to keep up to tabs, or not keep up, but the best way to help us grow. And in addition to that, we have a newsletter. It comes out twice a week.

It's really cool. The one on Friday is the weekly recap. The one on Wednesday is a thought piece covering a topic that will eventually cover on the podcast. And yeah, if you're new here, don't forget to subscribe. Thank you so much for joining with us. And yeah, we'll see you guys in the next one. See you guys.