WTF are Agent Loops and why are the Creators of OpenClaw and Claude Code talking about them?

I know everyone's talking about Fable 5 now, but we got to talk about loops as well because even you got people now writing about how you use loops for Fable 5 and it's just moving so quickly. And so I got my friend here, Matt Van Horn, who wrote an article on what the fuck are loops. And I think you got to have like 3 or 4 million Twitter views on it right now. Is it more than that or what?

Yeah, I think it's like 3.4 million, which is crazy. It's definitely the most views I've ever gotten on an article.

That's awesome. Yeah. So I wanna talk about how you made it go viral like that. But I think we should, you have a lot of people talking about, but I was just talking to my CTO about it on Tuesday. So today's Thursday. It started taking off Sunday when Peter Steinberger talked about it. Obviously Boris has been talking about how he's been using Loops and he doesn't prompt anymore. So you've got some good philosophies around it. I wanna talk about how you use the last 30 days as well. I wanna talk about your CLI library as well. And I wanna talk about anything else that you're interested in too, because I think there's a lot of higher leverage things in terms of the things that you typically talk about. But maybe you wanna, if you wanna introduce yourself first, I think that's a good starting point and then we can jump into what the F are loops.

Excellent. And I'm still not sure what a loop is, by the way, but I'm winging it like the rest of us. My name is Matt, based in Bellevue, Seattle area, and I'm a third-time entrepreneur. Haven't announced my new company yet, but we exist, just not yet. And previously I was one of the co-founders of the company that became Lyft, the transportation company. And then I started an AI appliance company called June where we put a camera in an oven and literally had an NVIDIA GPU on our Gen 1 June oven. Wall Street Journal called it the self-driving oven. And we sold that company to Weber, the grilling company, where I was an executive for the last few years.

That's awesome.

And now have a new unannounced company and have accidentally become a big open source nerd, which I am loving.

Love it. That's awesome. And we can nerd out on that too because I got my DG Arc Sparks living over here, but I don't maximize them. So we can talk about that too. So the conversation I had with my CGO is like, okay, are loops just spamming /goal? Because I've been spamming /goal for a while. And then I'm like, is it nuanced? Because I watched another video of someone else talking about it yesterday and I'm just like, I don't feel like anybody knows what it is right now. But all I know is, and you said this, You have the creator of OpenCLAW talking about it and you have the creator of Claude Code talking about it. And then now you have people, I think even Claude is even saying, oh, you should use Loops in Fable 5. So what is your definition of it? What's your take on this thing and how have you been using it?

Yeah, so what's crazy is that the quotes from Boris did not really go viral a couple weeks ago when it came out, when he was on Acquired. And then Peter just put out this tweet that he said it in such a way that sounded so obvious. Like, of course, no one's prompting anymore. Everyone just uses loops. And everyone's like, I don't understand what he just said. Like, there is a /loop command, there's /goal. And then there was this tweet that someone posted, and it was just like, no one knows what a loop is except Boris and Peter. And I was just like, yes, this is amazing. And so to me, what I think— what a loop is to me Is it's a cron job, so some sort of regular check-in, plus an LLM brain/judge that is kind of acting as if like a human used to do on a cron job, like, okay, this thing happened, they have to make a decision. Now it's like, okay, the LLM is going to make a decision within certain guardrails based on this cron job. And so there's different ways to call loops. Goal is is one of them. I have programs that I run while I sleep that I've been running for about maybe 4 or 5 months that run autonomously for 8 hours that I'm not using /loop, but it is an autonomous system that continues to run. And so that's, that's what I think a loop is.

I wanna come over to your, maybe your setup in a moment, but maybe like the, the people watching this right now that, you know, you got a marketing and, and business audience, so You know, people thinking, well, how do I make more money with loops? So how would you answer that question? Like, what would be some high-leverage examples of maybe doing that?

I have a very high-leverage example of not making money with a loop. So I put out a blog post this week. So someone tweeted that I told my agent to go make me $5 and it found a bug bounty and made $12. And so I literally saw that tweet. I was like, oh my goodness, like my autonomous system is designed perfectly for this. With the printing press, I can make a CLI for all the top bug bounty payouts. So I made this thing, called it the Bug Bounty Goat, and then I built this intricate system to autonomously, while I sleep, do research on different things to make money. So I built a money printer, and I ran this— I don't even know how many nights, but like over maybe 2 weeks or so— and I made $0. Like, literally $0. So I I feel like people just, like, there's this, like, hustle mode of, like, oh, I'm so rich, I made so much money. I literally failed and thought I was so clever and so smart and I made $0. So I have canceled that project and wrote a blog post about it.

Interesting.

But how to make money, I think, again, I think the biggest thing right now is the amount of effort it takes to build something that is decent is quite low. And I am not a software engineer. I have not made anything of value in software since high school until January of this year. And the fact that I'm able to build things that are decent right now is really interesting. And the thing that's missing is human taste, quality. And because the bar to being able to produce anything is getting lower and lower with every new model that comes out. And so the real question is, what do you make? What does create value? And so I think there's a lot of marketers and product people that have never been able to build before, that there's this ultimate test now. Oh, you say you're an idea guy or gal? Prove it. You now can. And so there's this test that could happen now that couldn't happen before. Yeah.

And Matt, I'll just share an example. I mean, on my end, 'cause this concept is so, it's so high-level theory for me. I just went to my Hermes and I'm just like, okay, well, Matt has this article over here, use the xAPI to read it. How do we apply this to our business? It's like, oh, well, you have one of our AEO SEO softwares called ClickFlow. It's like, you should just apply to the ClickFlow repo and look for bugs there and just keep looping it. You should apply to your Carrot repo and you should just keep looping it. I'm like, oh, those are obvious. I wasn't thinking about that. In a lot of these situations where I don't know what the heck is going on and this is so conceptual to me, I'll just take Matt's article here, which I'll pull it up right here just so everyone can see it. You got 3.4 million views. Oftentimes now, whoever's saying some smart things, I just copy the link. I just go to Hermes and I just say, how high leverage is this for us? What can we do here to help grow our business faster? And then you got this long-form piece over here, 3.4 million views. And this is a good starting point to make idea babies with.

Yeah, absolutely. One other thing that I'll add to it that's been really magical for me lately. I'm a big fan of this tool from Every called Compound Engineering that just writes really, really high-quality plans. And so what I'll do is I'll say, like, let's go with your example, CE plan, take a look at all— like, if an article comes out, like there was a sales article that my co-founder sent me that was about selling and a YouTube. And I was like, okay, watch this YouTube, grab the transcript. Watch this article, then pull up, use the Printing Press's Granola CLI. So Granola is a tool that transcribes conversations. Take the Granola CLI, now CE Plan, look at all my recent sales pitches that I've done. Literally you have all the transcripts and then run it through this framework of everything that I've done. It did that in about 7 minutes last night and gave me this really nice applied, here's the learnings from the YouTube, here's the learnings from the article, Applied to your exact sales pitch, which I've listened to 10 times in the last 1 week. And it was really, really great having that all come together to make it relevant for me. So I love taking full granola transcripts and putting them into Claude or into Codex and having it work with an article that you found and applying it to what's going on.

Got it.

So maybe if you can tell us a little bit, 'cause the way I do that right now, and I think everybody should be doing this, super smart. I think whether for us it's like, okay, we have the Granola API, we have the Gong API, which listens to all of our sales and customer calls. And then I just, because I have God mode with my Hermes, obviously my team can't do this because I don't want them to just YOLO with it. But since I have God mode, I'm just like, oh, can you do all those things? Can you make a battle card or can you find nuances like the customer objections and all that? Why should people be using— you created PrintingPress CLI, which I think is amazing because you can make everything into CLI. Can you explain what that is and what the benefit of that is, just so everyone can kind of understand?

Yeah, so, so I go to prettypress.dev to find what we have. The most valuable thing there is a library of skills of things that people have already printed. So someone made a Granola CLI. So if you use Granola, you definitely need to install that in Claude Code. You need to install it on your OpenCLaw, on your Hermes. Highly recommend it. And there's a lot of other things, like there's an ESPN CLI. CLI that sniffed and found the secret ESPN APIs that don't exist, but now they do. And so whenever there's— I have a lot of kids, I have 4 kids, and so I don't get to watch as many games as I want to, but if it's close at the end, I want to see that. And so for like the NBA Finals, I'm like, if it's within 5 points in the last 5 minutes, like I say to Hermes, use the ESPN CLI to poll it regularly and then only send me a message If it's close. And I've been doing this for every game. It's amazing. And but what's really magical about the printing press is most people are just using things in the library. But if there's a specific tool you have that there's no CLI for, what the printing press does is it, it runs this very serious, intricate flow to find the secret APIs, to look at the community, community CLIs and MCPs that people have made. And then it will create this very token-efficient and fast way for your agent to use that tool. So one of my friends, Chris Young, he's an entrepreneur that makes a hardware probe for your meat that has many, many sensors in it. And his inventory management company where he likes the warehouse where he stores stuff uses this very archaic old software. And so to look at his inventory, he has to log into this what feels like a 1990s web page. click around and it's just a horrible, painful experience. He used the printing press on this archaic software and found all the secret APIs, and now he's completely in Claude code, able to pull in all his inventory. He never has to go to that webpage again. It's really magical. And so anyone can print anything. Got it.

So just correct me if I'm wrong here. So the benefit of creating a CLI versus— because I'm not technical or an engineer. So The beauty of it is, one, you can maybe find APIs that you wouldn't have found, but two, it sounds like maybe it's faster as well. Is that— are those kind of— maybe they're about like weight?

The best way to think about it is if you've ever watched an agent try and fill out a form on a webpage, it's like, okay, hello, I'm now on alaskaairlines.com and I'm gonna book a flight. And it's like, you see the cursor moving, it's like, okay, where do flights get booked? Okay. Okay, great. Now I am going from Seattle to— right? And so it's just like clicking around aimlessly. The way to think about it is a CLI, a thoughtfully designed CLI— not all, there's lots of bad CLIs too— but a thoughtfully designed CLI is like Google Maps for your agent to be able to figure out how to navigate things. And one of the things we're working on at the Printing Press— we haven't announced this yet— is is self-learning CLIs. So it's the, the— so say you have the ESPN CLI on your Ermes and you say, okay, I want to know when the World Cup games are. But back when we printed the CLI, World Cup was not a thing. So who knows if it caught those original endpoints. But after it like brute forces finds them, it's actually going to leave notes for itself for next time that says, now I know how to find World Cup games. because I now have mapped this and left notes for itself.

Recursive self-improvement.

Absolutely.

Yeah, I love that. That's great. PrintingPress CLI, I think the cool things, I think you just shared Katherine. I think you retweeted her yesterday. She shared something cool. I forgot what it was. It looked cool. I just forgot what she shared. But people are doing some really interesting— can you share some other interesting use cases too? Because I wanted to play with it. It's just one of those, I just get bogged down by my Hermes.

So this, it's literally dead simple. So my, my number one use case on my Arise is literally for printing press related CLI. So what I would do is I would send it— there's a link on printingpress.dev. It's like copy and paste this to your agent. And it's, and it says, with everything you know about me, look at everything in the library and tell me which ones I should install.

Hmm.

And it will just do that for you. And what's, what's amazing is a lot of the printing press ones are better than the official ones. that people put out. One of the best examples of this, which has nothing to do with the printing press, but one of the inspirations for the printing press, is Peter Steinberger. He started making a lot of CLIs. So, he's made, I don't know, I wanna say like 12-ish CLIs for himself, for his agents, because he's such a power user. And one of the best ones he's made is the Google CLI. It's called Gog CLI. It's made by Peter.

Yep.

And months later, Google put out their official CLI.

Yeah.

And I said to Last 30 Days, hey, which one should I use? GOG CLI versus Google CLI? And I was like, of course Google's is going to win. They got to see Peter's and match it. And last 30 days was like, no, Peter's is way better for these reasons. On Reddit they talk about this, on X they talk about this. And that's one of the things that inspired the printing press. Like, why? How is this possible? How is this so thoughtful? And then I dug in and I got it and I was like, holy crap. And that's what inspired— making the printing press so that anyone can do this. So you don't need to wait till Peter's frustrated with an API before he has to make his own CLI. Anyone can print their own CLI and agents. And by the way, it doesn't just make a CLI, it also makes an agent skill. So if you use OpenCLAW, if you use Claude Code, if you use Codex, it makes a skill for each of these on top of the CLI to be controlled by it.

So check this out, I want to show this to everyone. So, and by the way, I didn't know that GOG was, uh, still better than GWS. Once it switched Once Google released theirs, I was like, oh, Google obviously would do a lot better. And you're telling me this and I didn't even know about it. And so I wanna talk about last 30 days in a second too, but I just wanna show everyone here. So here's why you should be creating other CLIs or using other CLIs for yourself. 'Cause you look at this over here, you can get nonstop flights over 8 hours from Seattle, December 24th to January 1st. So you don't have to go into a UI and search everything. You just type this into the CLI and you get this data over here. And can you actually book in here, Matt, too?

You cannot book. That would be nice. There are certain ones, like I've seen there's a Jimmy John's one, there's an order ahead one that you can book. Google Flights does not.

Okay, cool. When does OKC play next? Okay, great. To your point, and then we're talking about Rotten Tomatoes. A lot of good consumer use cases here, but from a business standpoint, I'm sure there's a lot of good use cases here too to pull what you need. It's just not— to your point, it's just a lot faster, I think.

Go down to that Contact Gov, and that's a fun one. A little bit right there. So, okay, so Contact Goat is cool. When we use the word goat, it's the greatest of all time. And it's pulling— it's a combo CLI that's pulling in multiple things.

Yeah.

So the Contact Goat uses a LinkedIn fan MCP as the first thing that it uses. LinkedIn has lots of bot protection. They do not want you to be able to use LinkedIn. So this tries to use your LinkedIn account. Then it uses Happenstance, which is a great service where you kind of download your whole LinkedIn database and it kind of recreates LinkedIn and allows you to pull stuff in. Then DeepLine is a service similar to Clay. It's kind of the CLI-based Clay that allows you to purchase verified email addresses and do all sorts of searches. And so the example I did in the printing press launch video when I used the Contact Goat, I was like, do I have a connection to the new Apple CEO? And LinkedIn told me that I did. It told me my friend who used to work at Apple. Has a connection to him. And I was like, okay, so I could get a warm intro, but also can I pay 10 cents and can you gimme his email address? And DeepLion paid the 10 cents and did this all within saying like, do I have a warm connection to him? Can you get me his email address? And so the Contact Goat kind of pulled all 3 of those together.

That's awesome. And by the way, like I, this is what you just mentioned, I think is the next step because what I had done with my previous OpenClaw I downloaded all my LinkedIn connections and I was just like, hey, who should I connect based on all the things that related to our products and services right now? I was like, well, you have this contact over here at Google. I hadn't talked to this guy in 10 years. Now this guy's high up at Google. He has a lot of these agency connections. He has a lot. And then what ends up happening is he invites me to speak at a Google conference and he gives me a lot of details on what we should be doing to do better for Google as an example. But that was a little bit of manual work. I still had to actively ask OpenClaw. I think in this case you can probably set a loop tied in with this to help you find these connections every single week, right?

Absolutely. I mean, that's a lot of the magic. I mean, one of the best loops, loop is not a word in my vocabulary that I need to work on. I call it automation or cron jobs. But one of the most valuable loops that I have in my life is running /last30days, which is another product that I made, which does this deep social agentic search where it checks X, Reddit, TikTok, YouTube. And so I have these, again, with my OpenClaws, how I had this set up originally was once a week it will run a cron job on the most important things in my life. So the people in my life that are important business-wise, Some of my investors, my competitors, being able to pull in these reports is so valuable.

This is great. So the way I see it, I'm like kind of how you're talking, really Loops right now as of today, June 11th, 2026, at least my understanding, it's a cron job. It's an upgraded version maybe of a cron job, an upgraded version maybe of like a /call. And is it, Matt, would you say it's on its way to recursive self-improvement? Maybe that's the next step.

Yes. I mean, the— what's interesting is Claude is shipping so many things and Codex team is shipping so many things so quickly, but we still haven't had this what I would call Ermes or OpenCLAW-like loop that you can continue that. So even when I type loop into Claude code, it's like, hey, when you close this context window, I'm going to forget about this loop. And like, that makes me sad today. But I feel like that's gonna be a solved problem in the next 45 days. Like, there's just no way that there can't be a thoughtful, clever way to do that. And by the way, you pulled up printing press marketing. So these are CLIs that the community has printed. So Catherine, who you just mentioned, has printed a lot of these. She's very deep in the marketing community, and so she's automated a lot of these. Trevin, who I built the printing press with, has email octopus. I'd never even heard of email octopus. And it's there. Dub is here. And so people are printing and making these things. Like, I don't know what Bento is. I don't know who that person is. Like, the community is just building things and then submitting it to everyone, which is really nice. This is great.

What I'll just say about these printing press marketing CLIs, like, these will actually make you more money because if you can just pull your Ahrefs data and then cross-reference it with maybe other email data that you have in Beehive and you can pull that data faster, And then again, you can cross-reference it with Google Ads, Google Search Console, all these things. Then what ends up happening is you end up having these strategies that are almost like these idea babies that combine together. And then you can just execute a lot faster and have your, whether it's your Open Claw or your Hermes, we call ours Hermes, but whatever you want to call it, I just think this is great. And here's the thing. My question for you, Matt, obviously there's a very big SEO play here because if you have it be like a Almost like a CLI site where people can review CLIs and then they can share their own. Is that kind of the vision for this? Because I can see this growing a lot organically.

Yeah, I am not an SEO person, but Trevin redesigned these pages to be thoughtful and mindful of SEO recently. But we don't have reviews right now. But absolutely, I mean, one of the things that we take great pride in is like, who printed this? Like, we like to keep that right and center. People should be proud that they added this to the printing press library and it keeps growing.

Matt, let me give you my unsolicited thought here. You can take it if you want it or not, but—

Please.

Honestly, I haven't tried this yet, but it's like what I would do here is like, okay, well, who's a really good directory site? You have G2, obviously. You have clutch.co for agencies and you have a lot of these directories out there. It's like, okay, well, now that we have Fable 5, can we just analyze those sites and figure out how to apply it to this site over here? And what would Fable 5 do exactly? And then That to me is a good usage of Fable 5 and perhaps a, I don't know if it's like a, maybe it's an ongoing loop, but I think there's an initial audit that it can do here because to me, you have a seedling of something that can drive a ton of organic traffic, whether it's AEO or SEO or whatever we wanna call it.

Yeah, no, thank you. And one of the things I wanna do is for a lot of these, I've gotten really obsessed with making launch videos on X. using Hyperframes. Before that, I used Remotion. And what I would love to— and so, for a lot of these CLIs, when we announce them on X, we put out a video. And kind of the automation is getting so good using Hyperframes that I would love to put a video demo on every single CLI on that page. I feel like would be really nice. Kind of a looping GIF.

And it's— I'm looking. So, okay. I found your video over here. I'm gonna share it with everyone because I think everyone should see this. And we should talk about how you marketed this, too. Because I think it's really smart. And if you can do this, obviously it sounds like your plan is maybe to do it a little more automated. That'd be cool too. So let's check this out real quick.

Yo, this one go out to every agent on the grind, hunting through docs in the middle of the night. The printing press in the building. Yeah, yeah, yeah. Ain't nothing more valuable than your time and your dime. In a world full of agents, that's tokens in time. Every wrong turn, every dot that you scroll, every wasted call is a tax on the whole. A well-designed CLI is muscle memory. See, no hunting, no wrong turns, no fees on the feet. So we built the print and press, the Print the best CLI's in the world for the agents on the quest. You point it at a spec, you point it at a site, hard file in your pocket, it'll print it overnight. Freeze the official docs, every endpoint in sight, then it studies every CLI. Let the community rock every MCP server, every skill in the night. Sniffs the web for APIs that were hiding from the light. Google Flights, Domino's, Kayak, that's right. The press finds the secret. Then it builds it tight, print and press. C-L-I, C-L-I.

I wanted everyone to hear the payoff of the chorus there. But if we look at this one, hold on, before I turn it over to you, this post has, it looks like 1.3 million views on it and you used hyperframes for this, is this correct? Yep.

Yep.

Okay. So do you want to talk? Because dude, If every one of your videos does even 10% of the views, I think it's a success. So walk us through kind of the thinking behind this and how you made it.

Yeah, so a lot of my videos, I, again, I'm not used to saying the word loop. I'll say it differently. Whenever I do something once that I feel like was too hard to do, too manual to do, I always make a skill out of it. I use compound engineering in Cloud Code. That's where I live primarily. And I say, CE plan, making this video sucked. Like, you saw how many times you let me down. The script made no sense. The font sizes were too small. I wanna make videos more often. Please make a skill that will be called MVH, make video, like whatever it is. And make a plan for that, then go execute on that plan, and then go ship the Claude code skill to my private GitHub repo for my personal skills so that I could use it next time. And so though that's the most valuable loop that I have, is whenever I've done something that I want to do again, to leave notes for myself and just run it again and then give feedback the next time, and then it gets tighter and tighter and tighter until it can start to become automated. And so a lot of my articles that I'm writing now are using this skill that I've now published. And it gets data because it pulls in what is popular and what does well. And so it looks for those similar patterns. And when it writes a new article with me and just kind of asks for the high level, it's able to get really, really smart and much, much better now.

How do you keep track of all your skills?

I just have one skill and I sometimes forget about the things I've added to it. Like, I've tried to build the same thing a second time and my agent's like, you already built that, sir. I'm like, oh, well, is it any good? Yeah, it's pretty good. Great. All right, run it and see how it does and then I'll make it better. So that, that does, that does happen. But just having— but again, don't keep it local. Make sure you publish to GitHub, make it private, and just Constantly be adding to it to automate anything you do.

And when you say you have one skill, because obviously you have a last 30-day skill, maybe there's a Skillify skill. I have a YouTube long-form skill. There's an X skill. You're saying you have one for each thing, right? My whole thing is when I ask that question, I tend to forget the skills that I've created. Because of that, it means I forget when to invoke it as well, right? And so it's like knowing I know I have the spell somewhere, but I forget How to invoke it.

So.

Yeah. So, okay. So, let me— let me— okay. So, anything that I will release publicly. So, I have— I think I have 3 projects that are public that are open source that have some popularity. So, I have Last 30 Days, which we should talk about in a second because that's my most popular one. We just hit 40,000 GitHub stars, which is crazy. We have the printing press. And then the third is something called Agent Cookie, which if you have a Hermes or OpenCLaw agent, having your Mac cookies sync perfectly with your agent in the cloud so it's always logged in. So those are my 3 public open source projects. Then this one that I was talking about, I'm calling it the, the MVH skill, is a private one. That's where I keep everything that's just for me. So that's where I have my autonomous do work while I sleep skills. That's where I have my content article creation. my video content creation. And so those are private ones that I have for me. And so it's just one skill that's becoming bloated and bigger and bigger, but it works. And so kind of I just keep that one and just keep adding to it. I just add a new command. And, and one of the things that I highly recommend when creating a skill is for yourself is use a really high-quality skill as your baseline. So for me, I'm a big compound engineering fan, and so when I created my skill, my MVH skill, I said, CE Plan, look at compound engineering and how it works. I want you to create an MVH skill, design your architecture after how CE compound engineering works. And so it didn't just make up how to make a good skill, it actually based it off one of the best and most thoughtfully designed skills in the world. And so luckily I got that for free because it's open source. So thank you. Thank you, Kieran. Thank you, Trevin, for Compound Engineering.

Got it. Love it. Okay. That makes a lot more sense to me. Okay, cool. So let's move over to last 30 days. So this became really popular. I think I started seeing it like a couple months ago or so. So can you explain what it is? Because I think this was one of the bases of you also creating this WTF are loops article.

Yes, absolutely. So Last 30 Days is an agentic search engine. And so what it does is it pulls in social signal from the last 30 days. So from TikTok, from Reddit, from X, from YouTube, from, and all sorts of other sources. We even, people have added, it's quite popular in China. People have added Chinese social sources to Last 30 Days. And what it does is it pulls in, it, you run it locally either on your OpenClaw, your Ermes, your your Claude code, your Codex, and you put in your API keys. And so, what that does is there's no place on the planet that has all the licensing rights to all these services. So, if you want to search X, great. You better use Groq CLI.

Mm-hmm.

If you want to use YouTube, better use Gemini. You want to use Reddit? Better use ChatGPT. They have a partnership with Reddit. Claude doesn't have access to any of these. And what Last 30 Days does is you plug in your keys and it's able to run the most thorough social search in the world on a particular topic. And what it does, it does a few things really well. So the summary is amazing. So being able to kind of take all that in and say like, here's what's going on in the last 30 days. Like, it's crazy. Like, what's crazy is I had, an investor of mine yesterday, uh, was using the last 30 days. He's like, hey, can you help me diligence this company? And I'm like, sure. And so I, um, he said, they said their video got a ton of views, but I don't see it. And so I was like, I don't know. So I just literally typed in the name of the company and I found it. And what's crazy, I didn't mention the video at all. And it was like, uh, this looks like a paid UGC campaign just on TikTok across 10 accounts That this does not feel organic. And I literally didn't ask about the video. It literally found that social signal just analyzing it on TikTok, which is crazy, which is crazy. And so it's, it's a really magical experience, and people have loved it. We, we were the top number one trending repo for the last week on all of GitHub. It just hit 40,000 stars. We're about to hit it today. Maybe, maybe someone listening right now, please star. I think we're like a couple hundred away from hitting 40,000 stars, which is absolutely wild. And this, this is something I built on the ski mountain. Our snow was really bad here in January.

There we go.

No, we need 400, please.

There's enough for this webinar, guys. There's a lot of people watching.

Last 30 Days skill, please, please star on GitHub. And I built this on the ski mountain because it's something that I just wanted for myself. And I made the video on that day and launched it, and I didn't know if anyone would use it. And again, got a million views on the launch video for the last 30 days, and it worked out really well.

By the way, there's multiple marketing things here, right? So one, if you're using HyperFrame, okay, you can make these videos that are gonna get a lot of views if you've made something novel, right? And obviously anything that's useful in AI right now, I'd say is pretty novel. That's one piece. But also to me, GitHub is a way of marketing too because for some of our repos, we have a call to action at the end of each page and you can do that. I think because agents are going to be pulling from these repos all the time, I just think it's a nice distribution channel too. I think the videos are nice for humans to watch and I think the repos are good for both humans and agents. Absolutely. So by the way, uh, it looks like you're— it looks like Trevin here is actually hanging out. He's actually answering people in the comments here that I'm seeing. So he's dropping links.

Uh-oh, uh-oh. What Trevin said? I don't see anything. I'm just in the, the Riverside.

Yeah, no, it's good. He's just— he's, uh, he's supporting people on, uh, on how— where PrintingPress, uh, is and like how it works and all that. So, uh, okay, so last 30 days is great. And by the way, this is— guys, this is all open source. This is free. Um, it's been great for me. Like, I use it When I'm like, today I'm recording the Marketing School podcast later and I'm like, oh, just gimme the last 30 days on the latest marketing topics and gimme strong hooks as well. I'll just go to Hermes and ask for that. So this is great. And so you mentioned, okay, there's the last 30 days, there's the printing press you have, and I think you just mentioned another one that we should talk about that sounds like I'm not aware of it. So what was the third one?

It's only been out for a week. You haven't been paying attention to everything I tweet, Eric. I know it's—

Sorry.

I'd not expect you to. It's, um, but I, I don't want to be a, uh, I don't know, too obnoxious where I'm like retweeting my same thing like 3 times a day. I'm like, okay, if I have something new and high value, like a new video, like I can post that. But I, I try not to be that person that is just like, have you installed Agent Cookie today? Like 3 times a day. I, I posted about it a few days ago. Yeah, so I'm gonna look for it.

You talk about it, go ahead.

Okay, so Agent Cookie. So if— so it's again, it's a niche market, but it might be your type of market. So if you have a Mac laptop or Mac main computer that you use and you have a Mac Mini that is hosting your Hermes or your OpenClaw, this tool is extremely valuable for you. For anyone else, it doesn't matter at all. Okay, and so what it does is it securely over Tailscale Takes your cookies, like your Chrome logged-in cookies. If you're logged into Instacart, you're logged into Amazon, you're logged into GitHub, and it keeps them perfectly in sync with your Mac Mini. And so if you like— so one of the reasons I built this is I have a lot of kids, 4 kids. We use a lot of Instacart. And so often on the go, I need to add things to Instacart because or else the kids are not gonna eat. It's usually adding to Costco. Big Costco family. And so I built, using the printing press, I built the Instacart CLI. But I never use my Mac Mini as a computer. Like, it's just in a closet at the office. And like, I'm not logged into instacart.com. And so like, am I supposed to remote control in? Am I supposed to like attach a monitor and a keyboard? Like, it's just like, I don't want to do these things. And so I was like, I wonder if I could magically fix this with a cookie sink. And it didn't work at first, but I kept saying, do better, do better, make no mistakes.

You can do it.

And, uh, you can do it. And then suddenly it worked. And so now, just because I'm logged into instacart.com on my laptop, my Ermes agent in the closet is also logged in. So I can just be like, hey, please, please add mangoes to Costco, and it just can automagically using the printing press CLI Plus the automagical cookies from Agent Cookie.

Can I ask a stupid question? One thing that pisses me off about my— whether it's my Hermes or whether it's my OpenClaw, is that it continually sends me .md files that it stores locally despite the fact that I'm like, bro, I can't click this in Slack and I can't share this with people. Stop pissing me off. I'm like, let's fix this permanently. It's like, it's fixed. It's fixed. I'm like, no, it's not fixed. It's not fixed. You keep doing the same thing over and over. So does this solve that?

I don't know why it's sending you .md files. It never sends me .md files. But I will say that you could set it up. Have you seen Proof by Every? P-R-O-O-F. So Proof is a tool that came out of Every. And what it is, is it's the best way to share .md files with humans. So a human should never read a Markdown ugly .md file. But if you do wanna share something with a colleague of yours, use proof editor and it will work really well. So for example, you could tell your agent, do not ever send a .md here, always make a proof file and post that to Slack. And if you click that, that'll actually format nicely, things will be bolded, italicized, And so I think that tool could probably improve your quality of life here.

Yeah, I think there's a couple things going on with mine at least. One, I think I've overloaded the memory with it, so it's starting to just forget things. But what it was doing was it was putting it into Obsidian and now it's just storing the files locally. But then we also make HTML artifacts too that I share with the team because again, to your point, I don't want them to read everything. But I think I'm gonna try Agent Cookie here. I'm gonna give it a star. It has 392. I'm gonna be 393 over here.

All right. Thank you.

Closer to 400 now, but this sounds really useful. I think, by the way, I really use mine mainly for business use cases, but can you give maybe some more family use cases? Because I've seen some people talk about it for like, it's kind of their family agent, right? So how do you think about it there?

Like how your OpenCLAW, your Hermes, Claude Code can be used for real life?

Claude Code for real life, right? Because I don't think about that. I just think about business the whole time.

Yeah, I mean, one of my best experiences I ever had was planning a Disney World trip in Claude Code. And how I did that was I said, last 30 days, Disney World. So that was the first prompt. So it kind of spit out and it was like, okay, you know, Thunder Mountain's gonna be closed. These new things are open. There's a new fireworks show that they're testing, but no— Oh, sorry, this is Universal. Except I ran it for both. Universal has never announced a fireworks show at the new park, but they've been testing it. So this is going. So it was kind of like getting like the Reddit inside scoop, et cetera. So ran last 30 days first, but I actually didn't even wanna read that report. I just wanted my agent to turn it into something useful. And then I was like, okay, CE plan. So it already had the context of last 30 days. And I was like, CE plan, I'm gonna be in Disney World these days. I'm gonna be doing one day in the park. Ideally I do, I'm a Disney maximalist. I'm a little crazy. I was like, I wanna do, 3 or 4 parks in 1 day, and I want to do, uh, have 1 day before a flight of the new Epic Universe Universal.

Crazy.

These are the ages of my kids. I want you to know which kids will be allowed on what rides. Please guess their heights. I don't know their heights. Like, guess based on their age what they're likely to be, or if they're borderline. And please create a thoughtfully designed itinerary for this. And, and I know Disney has this complicated Genie Plus, skip the line, complicated thing. And some of it you have to sign up for like a couple days in advance or a couple weeks. I need to know exactly when those things go on sale. And then I want you to tell me what to tell my OpenClaw to do to set reminders and add to my calendar when those things are. And I was like, CE work, go. And it just like did like 20 minutes of work.

Yeah.

Came with everything. told me which rides, a recommended itinerary, and then added to my calendar, set up reminders so my Open Claw would send me a notification that's like, hey, it is 5:00 AM, you need to register right now if you wanna get on this ride that you get to pay $20 to ride this one ride at Disney. And it crushed it.

I love it. And are you using— I'm just curious, so, you know, what's like your CLI tool of choice?

My CLI to— oh, like for my Claude code?

Yeah.

Yeah, I, I live in Claude code.

Okay.

And, uh, but I use Codex a ton, but I, I kind of use Claude code as the, the brain. And I know that brain word gets used so much, so maybe I'm not using it well, but— and Codex as the muscle. And the reason for that is pre-Fable 5, which just came out, pre-Fable 5, I felt like the models were pretty similar. I didn't think one was better at coding than the other, but I, I've written all my scripts and skills for Claude Code, but a lot of them send all the work to Codex and then sends it back because I don't run out of my Codex limits as quickly as I run out of my Claude Code limits.

Did you see the numbers on this? So a semi-analysis showed the other day, and because there was a month where I was just on Claude API and I, I was like, I spent $13,000 or something like that, right? I'm like, Whoa, okay, this needs to stop. SemiAnalysis ran a study. It's like for the $200 a month plan for OpenAI Pro, the 20x, you're basically for $200, you're getting $14,000 in usage. Then if you're paying for the same version of Claude Pro, you are getting $8,000 worth of usage, which is exactly the difference that we're feeling. I didn't know that.

Yeah, I did not know that. That feels Right and accurate.

Yeah. Okay. So, but, and then going back to it, you know, I think it's— I'm more so talking about that. So the harnesses I understand, but also the terminal app that you're using, because some people are like, oh, some people are just— I'm just using terminal. Some people are like, I'm using Ghosty. Some people are like, I'm using this. Like, are you using anything special there? I'm just curious.

So I recently switched from Ghosty to Cmux. And I I was losing notifications in Ghosty. Like, I love the purity of a terminal. So, I do really like living in a terminal-like experience.

Yep.

But there were certain notifications that I was just— I found myself aimlessly hitting, like, Command 1, Command 2, Command 3, just, like, hoping to, like, find the thing that needed human input when I've got, like, 8 windows going at the same time. And C-Mux nicely just moves it up to the top.

But it's still just a very And what do you say to the people, like, let's just say you have marketers, business people watching right now, like, why should they be learning how to use the terminal or using how to— learning how to use a CLI? Because they're probably comfortable in using just the Claude app or the Codex app.

Because the, the most valuable skill in the world is telling computers what to do. And to geek out and learn how to use Claude Code, to use Codex, to set up a Mac Mini, like these things, even though they might be painful for a lot of people. It's never been easier to brute force your way to success. Like, when I would— if— let's go— if there was an OpenClaw-like thing 10 years ago, I would have to find my smartest friend and be like, hey, I've got a stupid question. Okay, so it says I'm at 127.0.0.1 and the homepage isn't loading. And they're like— That literally means your computer's URL. Matt, you're a dumb dumb. And I'd be like, I don't understand, but that's the IP address, right? And now I can just screenshot this.

Yeah.

And put it in ChatGPT Desktop, put it in Claude Desktop and say, help, Eli5, I don't know what's going on. What should I type in here? Help me fix it, right? And so, for people that have never lived in the terminal, that have never— which I've never lived in the terminal before. I've been able to do basic things for the last 20 years, but never anything of value. But to get that initial setup going of just getting Claude code in the terminal, whatever terminal app you decide to use, just like you can brute force your way into it. Just copy and paste, copy and paste, copy and paste until it works. And I actually wrote a blog post that got— I think we got a little over a million views. And it's every agentic engineering hack I've learned so far. And so I highly recommend reading this article. And it is like the how-to from the beginning, step by step. These are the tools I use. This is how I prompt things. Here's how I use compound engineering. Here's how I use it for work. Here's how I use it for open source. Here's how I use it for Disney World. And it's kind of— here's how I copy granola logs into it. And so I highly recommend as a starting point, take a look at that article.

This is fascinating. I searched for it because I know it's an X article, right? But the first thing that shows is your LinkedIn one. But I guess it looks like you reposted LinkedIn, right? And how did it do on LinkedIn? I'm trying to see. I can't tell, but I know you got a million views on the X post. So guys, check out X, follow Matt Van Horn. I have a philosophical question for you and then kind of a wrap-up question for you here. So if you, the current Matt Van Horn, is competing with last year's Matt Van Horn without the superpowers now, do you think that version without the superpowers can catch up in a year if you keep compounding? Because some people think, oh, if I'm not on top of AI right now, I can catch up in a year. Do you think that's possible?

No. I mean, catch up, there's always going to be another level. And I think there was a a Reddit post talking about inequality based on cost too in AI, being able to pay more for intelligence and being able to— like, we had a childcare tiff going on maybe 2 years ago, and we were arguing over email, and I was using paid ChatGPT, and she was using free ChatGPT email back and forth, and like, no humans were involved in this pretty much. But like, it was like I had an expensive lawyer and she had like a state-assigned lawyer, right? So I think that there's gonna be inequality in paying for more premium AI, and we're starting to see that. And the other thing that is like, you just gotta get started now. Like, this is the most valuable skill in the world. Like, be the nerd, put in the time, build something, try building your own skill. It might break, it might not work, but you The tools are getting so much better at fixing these things. Like, I was deep in cursor land a year ago. And my apps sucked. They were worthless. Like, I could not— there was nothing for me to submit to the world of value. And then suddenly around Thanksgiving, when Opus 4.5 came out and the new codecs came out, it stopped being a toy and became something real. And I shipped last 30 days of January.

Yep.

And I literally, like, I remember my friend Matt submitted the first PR to Last 30 Days because he found a bug. And I was like, I don't know what a PR is. Like, I don't know how to do this. Do I just click accept? And I just, like, said, okay, approve. And then, like, someone else— his was good. And then someone else submitted one and it broke the whole thing. I'm like, ah, no, I need to learn this thing. But I did. And so I think the most valuable thing you can do is Be the nerd, be the student, ask your agent dumb questions, and just brute force your way into learning how to tell your computer to do crazy things.

Awesome. Well, Matt, this has been great. We should do this again sometime. And Trevin, thanks so much for sharing these articles too, as your kind of assistant moderator here. But—

Yeah, thank you, Trevin.

Yeah, what's the best way for people to find you online?

X, @mvanhoorn on X. M-V-A-N-H-O-R-N on X.

Easy enough.