Limitless Podcast - OpenClaw Overview: Which AI Agent Should You Use?

Episode Date: April 16, 2026

Let's explore the world of claws: tools that streamline digital tasks into a single interface. By comparing offerings from Anthropic, OpenAI, and Google, we determine which is the best for you to use.Highlights include the user-friendly OpenClaw and Claude for task automation, Meta's acquisition of Manus for AI marketing, and Perplexity’s model-agnostic approach.------🌌 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 Claws 1013:57 OpenAI5:16 Anthropic8:00 Meta Manus10:46 Perplexity14:00 Opensource14:26 Customizing OpenClaw18:04 Real-World Applications 19:10 AI Integration22:35 See Ya!------RESOURCESJosh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

You probably have six or seven apps on your phone that each do one thing. One controls the lights, a thermostat, your security cameras, each has its own login, its own interface, and its own notifications. It's horribly annoying. Now, imagine replacing that all with a single WhatsApp thread or a single text message through your phone. That is Claw, and it's not just for home automation. People are running Claws that monitor codebases and open pull requests when they sleep.

They even draft these investor reports by pulling live market data. It can do anything that you can do, but digitally on your computer. The word claw describes something different from the AI that most people know. A lot of people are used to chatbots, which you type into, and it forgets everything about you by the next session. And then there's agents that are capable of actually going on the web and doing things for you, acting on your behalf. But then there's claws, things that run in the background that can orchestrate complicated tasks that have persistent memory that never forget. And that is what every major AI company in the world is building right now. Anthropic, Open AI, Google, perplexity. Everybody is building a claw. And this episode we're going to walk through

all the options that you have, and which one is overall best for you? Yeah, the AI stack has changed quite aggressively over the last couple of months. If we think of like the LLM as like the brain, then Clores are kind of like the arms and the limbs that allow you to actually do things. Now, everyone's familiar with ChatGBT, GBT, and Claude. It's kind of like you ask it something and it gives you an answer in text, but it doesn't actually do anything for you. And then we had the progression to AI agents, which is basically strapping an LLM with a bunch of different tools, your email, Slack, whatever it might be. But it was still super clunky. It could like use the tool, but it didn't know when to use the tool. And that's

when OpenClaw and Clore started becoming really popular because it tied in persistent memory and context about you with these different tools to create a personal agent that actually worked. And that's why it went super viral. Remember, this was a vibe coded project OpenClaure, formerly known as Claudebot over the weekend by Pete Stey. And it went completely viral where it has now tens of millions of users building their own versions of personal agents every single day. So if Chad GBT was the brain, you now have open claw or all these different claws, which do stuff for you. So now that we're familiar with what a claw is, there are kind of two tiers of these claws, one being plug and play. These are companies like Anthropic, Open AI, meta, through their

Manist thing and perplexity. These are the ones where you probably have an account. It's $20 a month. They give you access to all the tools. And then there is the fully custom ones. This is what you are probably most familiar with being OpenClawe and all of the open source alternatives to the open-claw infrastructure. That's kind of for the hardcore tinkers or people who really want to build custom stacks for their claws. But we should be starting probably with the foundation labs, the labs that everyone's familiar with, because chances are you already use them and have an account open with them already. Yeah. So Open AI of all the foundation labs is kind of what took OpenClaw being a niche project to something more mainstream. They technically didn't acquire

Pete Stey and the OpenClaw project. It's technically still open source and they created a foundation, but OpenAI basically owns OpenClaught at this point. Pete Stey works. He's a full-time employee at OpenAI and he's building out their version of OpenCore for OpenAI.

And so in Sam Maltman's announcement post, he says he's joining Open AI to drive the generation of personal agents. And what's interesting about Pete Stey's take on this is he still wants to build open claw in the vision that he has for it. But something interesting that Pete Stey mentioned on his podcast with Lex Friedman is

one of the reason why Sam almost super excited about acquiring OpenClaw project as a whole and bringing it so close to ChatGPT is Sam realized a weakness in ChatGBTGT no matter how many apps or tools he created, no matter how much of an app store he built on top

of ChatGBTGPT, developers didn't really want to build on OpenAI because it didn't give them the customization capability to be able to build the types of apps that they wanted to do. And in their fairness, they can't envision exactly what AI apps look like right now. This is the Wild West at the moment. So he realized that it was probably best to build the number one

or bring the number one open source ecosystem for building customized personal AI agents into OpenAI. And that's what he did with Pete Stey. Yeah, so since the acquisition, OpenAI has built out a suite of different features which create their own version of OpenClaW for ChatGPT. A few come to mind. So they have computer use,

which allows ChatGPT to take over your browser or desktop itself and do a bunch of things. Now, it's important to know that Open AI to date has acquired a lot of memory and context about you, and now they're matching that with a personal agent, which is exactly what OpenClaw is. So that's great to see. They also have an AI browser called Atlas,

which allows you to do any of the browser extension type of stuff. So if you want to do research or if you want to browse your email or do anything that is in your online digital world, it can now also do that. And the final thing is, of course, their coding agent, Codex itself, which they integrate with the former two,

tools that I just mentioned, which allow them to do the building and creation of new things that you want. So you have this personal agent that sits within chat chiptee right now that can do so much more than just be a chatbot LLM. Now, perhaps you don't have an open AI subscription and you have something like Anthropic on Claude. Well, you're in luck. Anthropic has probably the most robust alternative to open claw. Oh yeah. They didn't acquire open clause, so therefore they were forced to build their own version. And they've deployed quite a few features and utilities that really make it a compelling offering for the average person who doesn't want to fully build a custom claw stack. And there's this whole laundry list of things, starting with Claude Co-Work,

which, funny enough, it only launched, what, less than three months ago? It hasn't even been around for that long, but it, two months ago. But it has been such a huge game changer because Co-Work works with your entire computer. It allows it to access files on your desktop, execute those files, and actually make changes on your home computer. So it can take over whatever you give it access to in a secure way. Just recently, yesterday, they released routines, which are basically cron jobs for those who don't know, they're scheduled tasks that you can do with your computer off. You'll find a lot of people have been hyping up purchasing Mac minis because they allow your computer to sit on your desk and stay always on. Well, these new routines

offer a cron job feature for your Claude instance without your computer needing to be on all the time. That's very cool. They recently rolled out computer use, which lets Cloud actually see your screen, move the mouse around. It can click, type, navigate apps, and it's all permission gated so it's very secure, you kind of choose what you wanted to have access to, and then it will go off and do whatever tasks that you ask it to do. And then finally, there's this little thing you may have heard of called Claude Code, which is probably the most popular coding developer engine that currently exists, all baked into one super app. So the Clod desktop app is a super app that is functioning as a hybrid claw that allows you to schedule these tasks,

manage with your computer, and do so whether it's turned on or off. It's a really powerful piece the software. And personally, this is what I find myself using the most. I love the Anthropic stack that they built, how there's connectors and plugins to a lot of the tools that I use that are third party. And I found it to be just very easy to set up, very robust and very secure in the sense that it only accesses the things that I explicitly give a permission to. Yeah, it's really funny. Two months ago, Anthropic was the most hated company from the open claw community for one simple reason. They started just banning people using Anthropic clawed tools for OpenClaw. And no one quite knew why. People thought that they fumbled by like, you know,

this was a missed PR opportunity. You could have gotten so many new subscriptions and users. But Anthropic wouldn't let loose why they were doing this thing. Fast forward two months. And they've released, I believe it's 30 plus features specifically to rebuild OpenClau within a sandboxed and closed environment. And their reason is pretty fair, which is openclore is kind of the wild west. There's a lot of security floors. It's kind of expensive. It's kind of hard to use. We create a packaged, curated, cheap version, all for 20 bucks a month for you to use, and you don't have to worry about losing all your photos or your hard drive information. You could just quietly and confidently use this product, and it just works.

And that's what they shipped over the last two months. That's kind of a little bit of a misdirection because really it's over the last three weeks. They shipped the bulk of those features. Just insane amount of delivery. It wouldn't surprise me if they're using some kind of closed open-core implementation that they built within themselves to actually build and launch these features. A few more things that are pretty cool about their version of OpenClau, which OpenClaught itself does not have.

You can currently shut your laptop that you're watching this video on right now and still operate their version of OpenClaught, because they have these things called dispatch and routines that you mentioned earlier, which allows you to do this. So they've really built a comprehensive suite that has rebuilt OpenClau in a much more meaningful way, and it's important because Anthropic currently has the, context and data on millions and millions and millions of new retail users, which is key. Like, when it comes down to it, if everyone's building an open claw, the key differentiator

ends up becoming context and data. And Anthropic is one of two major companies right now, the other one being Open AI, that has this advantage. So if you've heard all that, and Claude isn't quite for you, there's another competitor on the block that goes by the name of META. And now META, they actually haven't built their claw harness themselves. Instead, they acquired a company named Manus, if you remember, for about $2 billion in December, of last year. And Manus is this harness that deploys fully autonomous agents to write code,

deploy apps, browse websites, do all the things that you would expect in a claw, except doing so now with the meta operating system. So if you'll remember, we recorded an episode last week on meta's new local models. They are looking to converge these two models and the agent harness into one instance. That seems like it's on the right track to becoming a pretty serious open club competitor as well. But not for the reason that you might think. Manus and Meta have differentiated their open-claw offering in a very specific way. They haven't offered it necessary. The hit product hasn't necessarily been for retail users.

It's been for advertisers. Now, that's where Meta makes majority of their money through Facebook, Instagram, and the likes. They basically have a crazy advertising network. In fact, I read an article earlier this week, I think two days ago, where their advertising rev is now becoming extremely competitive with Google, which is like the monopoly on the sector for God knows how long. Manus specifically is being used to help advertisers not only create their advertising campaign, but select product imagery, control the tools that they use to kind of manufacture and articulate specific advertisements to a target audience.

It does all the research for them in terms of how much time they should expose an ad to a particular user. All these finer details is now managed by their version of OpenClaw Manus. And I think that's such an interesting niche because this is the sort of direction that Meta has started to take with their. LLMs as well. Last week, they launched their own foundational model, the first that they've launched in over a year called Muse Spark. And it's not amazing as a general LLM, but it's great for specific tasks. And they're doing the same with OpenCall. So if you're an advertiser watching this, or if you're someone that wants to market their product on any of META's apps, definitely because it's extremely useful for you right now and no one else is doing it. And the final claw in

a box service that we're going to be talking about today is Perplexity, who just recently released Perplexity Computer. And Perplexity Computer is interesting and slightly different. from the rest of these options because perplexity is model agnostic. It actually uses, I believe, 19 separate models that you can kind of choose from and it'll pick depending on what the task that you're using is. And that model agnostic allows it to do things that are kind of unique to the perplexity platform. It does everything that you expected to. It can do the research. It can design, code, deploy projects, whatever kind of miscellaneous tasks you want to do. But it also has the benefit of being able to do so. I mean, you could do it fully in the cloud on the servers,

but they're also going to rolling out a version that runs locally on your machine, similar to what we're talking about with Claude and OpenAI. So Perplexity is also in there. It seems like the most accessible one. I think a lot of people use Perplexity kind of as a Google extension, and the way they've integrated computer is very easy and simple to understand. In fact, one of the new things that I found was cool is they deployed a tax expert, where it will actually go through your filings and your W-2s,

and it'll give you the best tax advice that it can, being a non-CPA. But what I find interesting about this is just how accessible it is. I find that a lot of the other things that we've talked about, they're fairly open-ended. It's on you to figure out your use cases for what you want to use this claw architecture for. With perplexity, they really do a good job of kind of honing in on specific use cases for you, like the taxes, like recently they interrated with your bank accounts where I could actually see your specific finances, and give you these kind of unique but specific use cases. And I think for a lot of people, that may be interesting.

So I would say perplexity probably is the one. that holds your hand the most. And if you want something simple, model agnostic, that would be the one to choose. It's hilarious to me, because every time I think perplexity doesn't have a moat and they're going to die, they somehow, like prove me wrong. Like you mentioned, like, it's model agnostic. They sit on all of these other LLM providers. Their moat isn't really the fact that they can access all these different models. It's because they've created something called an agent harness, which is, I guess something that sits on top of all the agents that you would create, it helps provide the right context that the right time and using the right tools at the right time, which sounds like a claw,

but they just do it better across a variety of different models where Anthropics just focused on Claude and Open Air is just focused on chat GPT. And the advantage that they have is not only is it cheaper, but it's more contextually oriented around a specific task that you're doing. So you mentioned too, you mentioned taxes and you mentioned the financials, specifically like you can act like a CFO right now. That's a good example of like how it can go deep into a specific vertical for something that you don't want to do, whether it's taxes or sorting out your accounting or finances, and it can like figure that out in a couple of seconds. And all of it for accessible, I think it's like for like a $20 a month package. So very cool.

Yeah, that's the price of all these kind of all in one boxes. Is there about $20 a month to use the software that gives you access to all of these claw functionality tools? Now, that is the conclusion of the closed source all in a box, press one button and have everything done for you. Next is the open source movement. this is where OpenClaught lives and breathes. This is the most popular form of Claw currently, just because OpenClaught went so viral so quickly. And OpenClaught is, like everyone knows,

the open source version that started it all. It is the orchestration platform for agents that gets things done and is fully customizable. So if you find yourself as someone who is a tinker, someone who likes to experiment, someone who has resilience to fix a lot of bug issues that you may run into, this is the platform for you because of how customizable it is.

due to the open source nature, you have full platform control of everything that it does from the MD files which suggest how it should act and behave to the actual code base that it uses every day to actually make these decisions and enact on your behalf. If you're an open claw user, you're actively using OpenClaught or you've at least already tried it. It's more for the people who want the full customization, who wants to do anything that they can imagine with it. And there's also some interesting examples that you just you were telling me of things that people have done that you wouldn't really necessarily think is something an open claw instance would solve. Okay.

What's a problem that you might have five years from now? Let me explain. Let me paint this picture for you. You own a house, but it's missing one quintessential component. You don't have a pool in your backyard. Josh? Well, what if I could help you with that? What if not me could help you with that?

But what if my open claw or my claw agent could help you with that? Sounds lovely. Someone created an open claw agent. that finds 500,000 to $1.2 million homes without a pool, gets the images from Zillow or any other public website, renders a pool or an image of a pool into their backyard and then sends it to them with a cold DM or cold email, basically saying, hey, doesn't this pool look really nice on you?

You should like consider leasing us or using us to create and set up your pool for them. And the inbound has just been insane for it. So it scans satellite imagery for mid-market homes, filters lot by size, pulls homeowner direct from public records, and then gets in contact with them with the new rendered image. And the conversion rate, supposedly, from the claim, this is Twitter, this is X, after all.

There are some crazy claims, has apparently been pretty damn good. And if you look at the video over here on the right, it looks pretty photorealistic. It has like a video version of it as well, so you can see kind of like a pool floaty going around. Just a cool use of open call. And then this second example, which I found like super interesting, was this mother uses or currently runs 11 open claw agents to help her raise her kids.

Now, that sounds insanely dystopian, but the framework is, ironically, in my opinion, one of the most useful examples of using open claw agents ever. So one example is she has a job to maintain. So one version of using all these open claw agents is she only uses voice base. She doesn't type anything. Now, Josh, I know you're a massive fan of using voice notes or speaking to your AI alarms. so am I. It is just a much more efficient way to use it. So she doesn't code anymore. She just speaks to agents about the specific type of thing that she wants to build in her particular job and it goes away

and it does that. But then she also needs to educate her kids. Now she homeschools her kids and what she does is she takes her phone, she takes a picture of a syllabus that she wants to teach her kid for the day and it generates a personalized lesson plan for each of her different kids tailored to their specific preferences and she walks them through it. Now remember, it's homeschools. So you don't necessarily have all the apparatus in your home. She provides that context to all her open-claw agents, and it knows which tool to use. So, like, if there's a toy that she can use to demonstrate the specific lesson,

it prompts her to go grab that toy and do all of this. So you might be wondering, well, this is completely insane. Why would you use AI to teach your kids, blah, blah, blah, blah, blah, blah. The net-net effect is it's freed her up to be more of a parent and spend time with her kids. So all of that to tie a loop around and says she has more time. on her hands to spend time with the kids. So I thought that was a super cool use. Now, going back to your point around open source open claw agents in particular, the customization is like a massive win for

people, but it's also how cheap it is for you to be able to run it on a local device, maybe it's on your phone. And also, if you want to run like completely crazy projects like raising your kids using 12 open claw agents, you probably don't want to be spending a bunch of different money on it. And so she specifically uses pico claw, which I'm referencing right here. And I think in some instances, iron claw when she's using personal data and she wants to use like a rust-based architecture. All of that to say is all of these are free. You can download them, access it yourself, and you can customize it in whichever way you want. So if you listen to this and you come up with a crazy idea, please do it and let us know what you build because like we're trying to track all of these

use cases whenever we can. Yeah, and a testament to the claw architecture being so resilient and robust. The real limitation, as it's always been, is how you can most creatively implement these tools to be most effective for you. And the ceiling is still, like, largely untapped. We don't know where this clocks out because people are doing really cool and interesting things with every single day. There's so many great examples of it and great enhancements, too. Now, what sits on top of this layer even are the markdown file. So if anyone who's familiar with OpenClaw, there is like a sole.mfd file. And there are a series of these markdown files that kind of give the model intention. They tell it what to do. They tell it how to act. And there's an entire kind of

pseudo industry that's been built on top of this orchestrating of the agents of telling them how to act by embedding interesting markdown files with unique sets of directions. Now, Gary Tan, founder of Y Combinator, he appears to have built one that has kind of become the most popular one, perhaps, which is G-stack. Can you explain what G-stack is and why it's this kind of valuable second layer to be built on top of your clause? Yeah. It's, I'm I've seen conflicting opinions on both sides about what Caritan is built here because he hyped it up. Basically, he said, I have made open call into this insane personal system that actually does productive work for you. And mine's the best implementation of that.

A bunch of engineers then evaluated what he had actually built and realized that he had just wrote a very extensively, and in his fairness, very well articulated, meticulously detailed document, which describes what he wants his agent to do. And, you know, Garretan, if any of you have watched any of his interviews, is an extremely verbose kind of guy. So it makes sense that, you know, he's worked with a bunch of these companies. He's had a bit of experience. He kind of knows what to ask for. And so the point that he proves with G-Sack is, number one, here's an easy and simple way that you can level up your open-claw agent in a matter of minutes. Use his framework and architecture and just fill in the blanks, like add in the details for your specific project.

And you now have an amazing blueprint to build an amazing open-core agent. But number two, how easy is it that it just requires two documents that you can spin up on your MacBook or on your laptop right now that can change your agent for the better? So the criticism for a bunch of industry folk was like, Gary, this is amazing, but this isn't anything necessarily novel. It's a skill issue. You need to be able to articulate. And that's basically something that is very apparent in the AI age. It's moved from being able to know how to code to be able to know how to communicate and explain. And that's the trend that we see with OpenClawn and a bunch of these other AIL alarms. And thankfully, all of this explaining happens in plain English.

And I think that's the most exciting and empowering part is it doesn't require you to be technical or to be a developer in order to improve these things. You just need to speak to it and kind of coerce it into doing the things that you wanted to do, whether it be through Anthropic and Claude or through the OpenClaw instance that is fully open source. There is so much customization to be had. And I think it's kind of on you to decide where you fall on the spectrum of plug and play. where you just kind of want a claw on rails that's something like Anthropic can provide to you, or if you're the tinkering type who really wants to build custom stacks like that mother who is using it to help her parents or a company that is being built around creating pools. There is no limit to the use cases you can find.

It's just a matter of your preferences and how you want to engage with this. But it very much seems like this is the new architecture that's going to stick around. This is kind of the foundation of what an AI first operating system could look like. and it's very clear that all these companies are converging on the same conclusion that this is true, and this is a very important part of the stack to own. And I think the one thing that is for sure is we are going to see this continued improvement and progress on this at a very rapid rate. I mean, Anthropic, they've been on fire, man. They've been releasing new features for this every single day at this point.

So I encourage anyone with an account anywhere, really, just to try it out and start flexing that muscle to see where it can improve these whatever parts of your life. Yeah, I wonder where this is going over the next couple of years. You have all these companies spending hundreds of millions of dollars to either acquire the thing or build the thing. They're all kind of building similar products. I think what it's ultimately going to come down to is the data and context that you can use to train these things. After all, that's what made it a differentiated product. And that's what makes it an exceptionally good AI operating system. Two companies that I've noticed haven't really got a version of this,

but probably should is Google. Well, Google actually does. It's called Mariner, right? But no one's ever heard of that. Right? No, but no one seems to use that. And they seem to have kind of left that by the wayside to focus on their main LLM.

And the other is Apple, who I'm actually more optimistic on because I know that they've signed the deal with Google Gemini and they're going to be launching AI Siri, hopefully for the fifth time on a failed attempt later this year, which should end up becoming a really useful version

of their open, agent. So I'm excited for these companies to launch their things, but I believe that is all we have for this episode. As Josh mentioned earlier, if you are listened to this and if you are inspired by either the closed sandbox version from Anthropic or the open source version from a picor or an iron claw, please, we encourage you to go out and use these things because there is a draw of viable use cases for a few these different AI things. Everyone has ideas, but no one's actually building the thing. It's never been easier to vibe code a custom AI agent. And we want to hear what you guys are interested in and are building. If you are a listener on this show and you aren't

subscribed, please do so. If you're listening to this on Spotify or Apple Podcasts, please give us a rating. Leave us a comment. We always love hearing feedback from you guys. And we'll be back tomorrow for the weekly roundup and possibly a surprise. There's a rumor in the mill that Opus 4.7 is coming. So we may have some big news dropping later this week. We'll see, but we will be here to monitor the situation as always. And as always, also, thank you so much for listening, for watching wherever you are. And we'll see you guys in the next one. Thank you.