Top Repos Clearly Explained

0:00Today we're going to look at the top 10 repos on GitHub and X. This is what's trending and frankly I'm not a developer. I don't usually spend time in GitHub, but I've got the person who taught me how to vibe code, how to code with Claude code, and he's going to explain to me what I can use from this and frankly what it all means. Presented

0:17by Zapier, the AI automation company. Let's go into number one. Adam, context mode, explain the problem because this one seems like it's useful and then you told me actually I'm not going to recommend it. What is this? Yeah, I I you know, often these MCP tools I I think you've used a lot of them, other people have. We're just installing them willy-nilly though, right? I I want Claude to talk to my Gmail, to talk to my calendar. You install that and then when you ask Claude, hey search my calendar, the calendar MCP tool will send back the whole month's worth of data or it'll send back huge threads of email. When

0:53you install them, then you ask a question, what am I going to work on tomorrow? Well, what happens is Claude will pull all of your events for tomorrow, maybe all the events for the next month. It's going to load a bunch of Jira tickets. It's going to check your email and all of that gets shoved into the context of the AI and then the AI is like, hold on, that's way too much and it kind of gives you the answer that you want, but man it's slow, it's expensive, tokens are getting more expensive. So you're sitting there waiting and a tool like this, what it does is it sits between your AI and

1:23these MCP connectors and it says, okay, I know that that Claude or that that Gmail sent back all of these emails, but only these three should go to the to the AI. Oh, you sent back a whole month's worth of events, only tomorrow's events should go to the AI. So it's pruning all of that out to keep the AI fast,

1:41efficient and cheap. So why didn't you get as excited about it as I thought you would? This seems like something I should be installing. This is a problem that Anthropic wants to solve, that OpenAI wants to solve. Like they're already 6 months ago they were terrible at pruning the results coming back from MCP tools. They're already 10 times better and 6 months from now the models themselves are going to be better at this. So maybe this is a short-term band-aid, but if it caused me to have to change my workflow or learn anything new, like I don't I don't see this as a as a durable tool. Meanwhile

2:14on Twitter people are getting psyched about it and I've got a link to it. Stop paying for Claude Max before you upgrade your plan. This is not going to keep you Adam or frankly even me from leveling up to Claude Max. We still need that that much usage. Okay, all right. And of course for everybody we've got this whole report including links directly to the to the repos available in the show notes. Let's

2:35go on to number two. Omi, this one I kind of added on cuz it was popular and I just don't think it's really really useful for me, but let me show you why it got popular. This is the founder's tweet. It went viral because of this. Let's see, it's loading up. [snorts] Here, I'm going to I'm going to just do

2:54a play-by-play. This guy is now looking at the internet and as you're seeing some dude drive, it looks like it's Epstein and then Omi says, no, that's not Epstein. As he's doing his work, Omi's watching his screen, guiding him, telling him not to be such a wuss with his boss, don't apologize. He then gets a call from his from Maya, his significant other and as he's talking to her and getting excited about her, he wants to walk over and show her some artwork that he got for her and then if you look in the background for a split second, there is a dude in a towel behind her. Her eyes

3:26light up. The dude gets away, she angles the camera away and our guy comes back with the picture that he's showing her, that he's got for her. It's wonderful. Meanwhile he didn't know the other guy was there in the background of the of the phone call, of the FaceTime, but Omi as it's talking to him about ordering food, Omi says, should I order for two people? And he goes, why? And he goes, well, for the other person in the

3:49background. Let me show you a picture. Now this is the way it's supposed to work. It's not exactly there, but this is why it blew up. This plus I installed it. It is creepy as you can imagine Adam to install this thing. Because [snorts] as soon as you install this app on your computer, it wants access to everything, otherwise it doesn't work. It needs to see my screen, needs to see to hear what I'm

4:11saying, needs to all this stuff. And I got really like anxious about it and immediately when I got anxious on the bottom right it said, go see our open go we're open source, click here to see it. And it linked over to to this. So I think the reason obviously that they open sourced it is to reassure people you can actually see how it works and feel better about it. You can use it to connect it to different devices, so it doesn't just have to be your

4:37computer. It's nothing that I'm necessarily going to be building with, am I right on this? But essentially this is somebody saying, look, trust me, you can look at all of our of Yeah, Andrew, I think this Omi here is is a great example of what's coming. Mhm. You know, we're going to see lots of these. We've seen a lot of little versions. Like this to me is just a more user-friendly version of Open Claw. You know, Open Claw you had to do all these integrations, but oh, now it's a second brain for my life and it can work while I'm not there. Like Omi looks like it's doing the same stuff, but in a

5:08much more user like look, there's glasses, there's a device you can hang and you know, they want to sell these devices and their AI service, but a year from now, 5 years from now, so many of these will exist. Like think all the way back to like Clippy, the Microsoft, you know, paper clip that was trying to help you write your your resume or whatever in in the '90s. And now that's what these things are doing. So I'm really excited for these. Omi's probably or Omi is probably not for me, but I'm I'm generally very excited and even the little device there, like I have a little clip-on recording device

5:41just like theirs that I can I'll take on walks with that. It's this. It's the it's that one. It's right there. Yeah, that's the one. This is the Omi works with it. Yeah. device. So yeah, you know, I I I I think this is just clearly where it's going. This one might not win, the others today might not win, but this is the future. I think open sourcing is going to help them and it'll help people be in their world and trust them more. Okay. Uh Claude context, Claude MCP that indexes your entire code base into vector DB, so Claude code can semantically search millions of lines without loading them

6:14into context. Again, this seems like a problem that here this is Sharbel. Sharbel loves to write about GitHub repos that are hot. Said you paste your question into Claude code, quote refactor the payment module to use the new retry logic. Claude hallucinates, it references files that don't exist, it misses the actual retry helper. Anyway, this is the problem that it's supposed to solve. Seems like a reasonable issue for me. Again, you

6:39weren't excited about it, right? this I mean it's a huge problem. I'm very excited about the problem. Explain the problem to me. I I don't know that I fully understand it. problem is, you know, you we we've talked about this context management and oh, how can I have Claude search my, you know, the 50 emails I got today and that's inefficient. Well, when you look at a code base, there there could be thousands of files, many of which are

7:00hundreds or thousands of lines long. There's a lot of code in all these applications. Well, how is Claude supposed to search that? If you're want to write some new function, you want to add a feature to your user interface, it probably has to update six files and it has to update them consistently and how does it even find those six files? So this is there to say, well, how can Claude effectively find the spots in

7:24your repository that you need to change? It's a real problem. I'm a little let's call it skeptical of this solution because the engineers that are that are here in the room, like we've been using tools like this for more than a decade. It's called the language server protocol and every development environment, VS code,

7:42cursor, whatever, they all support them. Many of them have them built in and the language server protocol is it's this. It's it's a way to track functions, to track files that relate to each other and you know, if I'm going to pick I'm trying to build something and you tell me I have to pick between this like likely vibe coded new AI driven whatever and the thing that's been working really well and that's been battle tested for 10, 15 years, like I'm going to go with the one that I know works, especially when it matters. The the problem with all this vibe coded stuff is that

8:16it's random. It's not it's not tested. We don't know when it works and there's lots of corner cases that that the language sort of server protocols have solved. So I'm using VS code to access Claude code. I don't need this. It's built in. If somebody's using terminal, would they need it? It's either built in. If you're

8:34using the terminal, you might need it. Um although if you're using the terminal, my guess is you're more of an engineering type anyway and you're already aware of language server protocols and and all the things there. If you're in a IDE, one of these developer environments like like VS code, it's either already installed or they're very easy to install through the extensions and you can ask Claude to install it. Let's say you're using a programming language that's not so popular, just ask your Claude, you know, the Claude tab in there, hey, can you install the language server protocol and any helpful developer tools? And it'll go and find all these like again battle

9:11tested tools that that already work. Okay, so real problem, not necessarily the right solution. By the way, this this conversation is sponsored by Zapier. Zapier's what allows me to bring in up to 10,000 other apps into whatever I'm building and that means they've got MCP, they've got SDK, they've got so many different ways to connect it. And what I love about it is I'm not giving the agent that I'm using complete control over something like Gmail. I'm limiting it and saying you can read and you can write drafts, but you cannot send email for me. I'm feeling completely safe. I also love that if I dump on one of these agents and move

9:46into another one, I just take my collection of tools from Zapier, just give just point the new tool at the Zapier connection and I'm done and it just works in a way that makes me feel safe, which means I get to try more things instead of feeling like, "Well, should I give it access to Gmail?" I

10:00can. All right. I'm adding Multica here because it is so insanely popular and because you're so frustrated with it this week. Do you want to like talk about what this is and show it on your screen? I'll stop my screen share, okay? Yeah. Dude, I I will at the end of last week I said I I need to try it um and so

10:18I did. I'll share my screen here. So, what Multica is Mhm. is we've all got all these like chat sessions all over the place. I'll I'll drag over one of my I use T3 code often to manage um to manage my chat sessions. So, here I've got a bunch of different projects. Uh each one has a bunch of chats. There's even more in all of these. And these are applications that we're building. They're tools or they're research projects. And I treat each one of these chats more or less like a ticket, like a you know, I'm trying to develop

10:54Say that again. Why Why are you using that instead of like VS Code or something else? What is that thing that Well, when you're in VS Code, you're you're looking at one project at a time and and I'm working on six projects at a time. And so here in T3 code, here's one project. Here's a totally unrelated project, unrelated project, unrelated

11:09project. is this So, is this So, is this another way for you to interact with Claude code or whatever agent you're using? Yeah, when I when you when you drop in, you can select and I'm switching regularly between both Claude code and using these models or the Codex uh harness and models. Should I be using this? I think it's really nice. I you know, it it's a it's a very simple app. It's a very You've already seen all of the features of the app. There's not really file editing, there's not previews. So, it's nice for someone like me that has some very specific problem. I want to manage

11:41a bunch of these and chat with them all. However, the problem Andrew when when you're here is oh well, there's five chats that are running here, there's two up here, and I really wish that I could treat these more like tickets, right? Like you know, I've talked earlier about solve problems. Hey man, we've solved how to organize work across a bunch of different projects and people. And it often looks like this, right? It's Trello, it's Jira, and what's happening is you create a ticket, you define the work for the ticket, and then you assign it to a person. You know, in this case it's created by me and and you know, I

12:15could assign it to someone on my team. Well, what Multica does it says, "Yeah, that's awesome. You could assign tickets to people in Jira, but in Multica, you can assign tickets to AI." And so, in Multica, you can set up in your workspace a bunch of agents. Unfortunately, Multica, I think it's also kind of in maybe the prototype stage. All of my agents were deleted uh a couple days ago when I updated the app. So, I'm not going to update the app anymore, but here I built one back just as a a tiny example. This is a feature agent and you could have a debugger agent and a research agent. They're just

12:48prompts. You just define a prompt for this. You can give it skills like you would uh in Claude code or anywhere else. And then you can connect it with you know, whatever data you you want to give it. All right, you have your agents. What do you actually do next? You say I've got this file. Well, instead of assigning this to a person, you can actually assign it to an agent. So, I'm going to go ahead and

13:10assign this to an agent right now. And then it's paused because it's in the backlog, but if I move it into the to-do phase, what we'll start to see is the agent will say, "Oh, this is my job. I see I have a ticket now." Just like a chat message. You know, think of this is the message you would send to Claude

13:25code. But oh, now it's working. And I've actually set up internally here uh wherever it is. Repositories, no. No. So far, I'm liking this. I don't see a problem. I've set this up with Claude code, Codex, and and Cursor because those are already installed locally. And And look, like in the background, the agent's still working. It's awesome, right? Like this is sort of the dream. Like I I was so excited to try this. I was so excited to get in. The thing's working. And And for me, it's it's just as simple as all these chats, right? Cuz I could just think of every ticket as a chat. And

13:58what happens is once it finishes its work, it'll move this and say, "Okay, you need to review this now." Well, I can chat back. I can move this into the done. I want to show you one that I had already worked on just as a great sample. Create a nice read me for the project. Simple, concise, useful. Okay. And it's connected to a project I set up. This is a real project, Reach Studio. This is an application that we're building at Gateway X. And I ask it to make a a read me. Okay, it

14:28ran some AI. It walked back and forth. It had you know, just like you've seen with AI, it has a bunch of thinking. It changes some files. And then it says, "Awesome. I added a new read me that explains the current Reach Studio app." Well, I've got this this repository checked out on my computer. I went and looked. I was like, "Oh, I don't see a read me in here." So, I asked back, "How do I find this? Where do I look to see what you did?" "Oh,

14:51look at the repo root for readme.md." Okay, so I went back and looked. I must have missed it. Nope, I didn't miss it. So, I click on this. And oh no, this isn't a link to the read me file. This is a link to I don't know what. Some website. Okay, that's a little weird and that's what all of these are. They're clearly just like AI, you know, whatever links. This is readme.md the website. And then it's like

15:16Come on, man. Like Okay, I don't see it. I don't see it. Right, but like what? And when you start to see a lot of question marks, you know, I'm getting frustrated. There's no all caps yet, but it'll come out. Um I checked GitHub. You know, I went on GitHub. I was like, "Did you push this to GitHub? Is there a new branch?" Like I don't even know what the heck it did. Okay, I could We're talking about the local multi Multica workspace, not on GitHub. Here's where I checked

15:38out. I was never going to find this. This like random folder on my computer. I was never ever going to find this. And this whole time, it's assuming that Okay, that's weird. I added the read me there only. Okay, so now I've got to figure out how to open it. There's no link to open it. There's no folder. At this point, like well, why am I not just chatting over here? Like all all of this like nice stuff where I was hoping it could work in the background and move tickets through. Like look at how much hand-holding even a teeny tiny little change like add a read me file is. This

16:10agent is is not ready for prime time. Now, maybe you could get over this by spending a huge amount of time prompting in your agent instructions and building a bunch of custom agents. I had built some custom agents again before, but they're very unfocused. There's also not much set up like in the project. I can't actually tell it where I want to check out the Reach Studio code. So, in my opinion Andrew, this feels like a Figma prototype, you know, nice functional, cool demo that I'd use for fundraising,

16:40but not really ready for prime time yet. All right. I can see then why people want it. That dream is what I'm looking for, but I also I could see that if it's not working exactly right, it's it's just destroying work. Like you literally said that it deleted everything that you had done because you

16:58you updated it. So, if you find another one of these Andrew, I want to try it. This is an awesome idea. It's a really cool direction. And maybe in in a couple months, they'll also be a lot farther down. So, I'm really excited to follow what they do. Uh but I'm I'm not switching today. All right. Voice Box. I really like Voice Box. Last week, we talked about this, which is the ability to create voices without having to pay 11 Labs. And apparently it really it's it's super popular because

17:32it can duplicate voices really well. But to use this, you need to write code yourself. To use this, you need to embed it into projects you already have. What we're looking at this week is Voice Box. And the difference with Voice Box is this is an app you can actually install on your computer and start using different voices. And that's the exciting part. In fact, here we've got Voice Box.sh right here. And I'll unmute it so that people can hear what it looks like. And what we are using is now the Jarvis voice. And he's typing in, "Sir, I have completed the analysis." And now we're going to get to hear Jarvis, I

18:07hope. Sir, I have completed the analysis. Your code has 12 critical vulnerabilities. Your coffee is cold. And frankly, your commit messages could use some work. Now, obviously this sounds like a robot, but if you continue, you can click on someone like Morgan Freeman Free Morgan Freeman. Uh and so, he let's hear Morgan Freeman.

18:29Generating it. And it's generating. penguins, galaxies, and the entire history of mankind. But nothing prepared me for the moment a computer learned to do my job from a 5-second audio clip. Okay. And now if we look and when zoom in, it's using Qwen. This is a model that it is using to make this work. [snorts] I think this is really interesting and it's exciting for people like me. We have actually in the past four podcast

18:57episodes had to create somebody's voice. We're creating a short, we need to do it. I totally see the value of using this for me as as a user. I don't yet see it for something I would build on top of a fork or anything, right? Yeah, maybe. You know, my my first question when I was looking at this is, "Well,

19:14can I connect it to the 11 Labs models?" Like I I love what they've done. I and I like I'm so excited for all these local models. I want to download this. I'm going to play with it. I might use it, but then of course, yeah, how do I actually use it for you know, real work? Sometimes these local models, they're just not quite good enough. So, I'm very excited about it. And the local models are getting better and better. Our computers are getting better, too. So, this is this is really a cool one to watch uh and and and try out. If you do anything with voice today,

19:44um you know, you should be trying this. See what it's capable On GitHub, there's no way for me to search through this whole repo. I could I could go to a file, but I can't just search for a word within all these text files, which is essentially all this is, right? Yeah, sorry. Uh get GitHub is not known as like a reliable and good user

20:03interface. Wow, cuz I would like to actually do a search and see is 11 Labs a possibility. Here's what I did since I couldn't search here. I um I use Claude Co-work to just start flipping through cuz I was creating all these uh pages in Claude Co-work. And as far as we can tell, no, there is no 11 Labs, but it might be coming in the future. All right, let's

20:23move on. This is so controversial. Android reverse engineering skill. What is What is this? Yeah, you know, you've got all these apps. Like I I have an app on on my phone for my Kia, and the app is not good, as you could probably imagine. And so I'm sitting here and I'm thinking, well, man, I I wish that I could wrap some of these. I want to have a button on my desk or on my computer where I can turn on the air conditioner in my car remotely instead of having to

20:50log into this stupid app every time. Well, the app doesn't have documentation. Kia has not published the docs for their API. So, there's no way for me as a builder or even Claude to know how to integrate with with the app. So, I can't really ask Claude to fix that. What this is reverse engineering is saying, well, I don't need the documentation. If I just watch this tool work from every different direction, and I click buttons in it, and I monitor what happens over time, I'm going to get to learn how the app makes the API calls. And if I write all of that down, eventually I'll have enough information

21:27to reproduce some of those calls myself. So, that's reverse engineering, and that's a it's a this is an old skill, and honestly AI is very good at this skill. Some of the some of the drama about the Claude Mythos model that you might have seen is around how good it is at reverse engineering, which gets close to hacking, you know, like as you point out here like is this legal? Is this always legal? Is this against terms of service? Like like reverse engineering is is a you know, a white hat or a gray hat form of form of

21:59hacking. And that is why there's so many people who were were bothered by this. It's one of the reasons why there's so many people who are bothered by this. Here's one of many people who said this is scary when they saw it um because you can reverse engineer and find problems. In fact, people are using this to understand to here to do security

22:17research and audit software. Um the other thing that has led to pushback, here it is on Reddit, is people are basically saying you just kind of vibe-coded something and you're putting it in the world, and it's almost like he hadn't earned the right to share it, and there's a little bit of suspicion around this also, and and the person behind it. I think I

22:38think I'll add one other weird thing. I've been looking into him to try to get a sense of who he is, what he's up to. He has a lot of really good use cases, but I think there's like this aw shucks, I've just created this thing attitude that maybe is making people feel like he doesn't he hasn't earned the right to do this. And I I'm He has said that he used Claude Code to build this, and I think at times he's said that he's used it that it's taken him a few hours, and it feels like irresponsible use of vibe

23:05coding, I guess. That's where some of this drama comes up. Yeah, maybe. I I think one one spot where you're going to struggle if you try this today is is when Anthropic released their 4.7 series models, they really added in a lot of restrictions around hacking and what Anthropic considers unethical behavior. And I bet if you try these skills today with 4.7, you'd say, "Oh, awesome. Like it's a more powerful model." It won't let you do it. I've been I've been turned away in the last week by a lot of things from Anthropic where it says, "No, I won't do what you asked me to do." Some it's the

23:43first time I've seen that from AI. I've seen AI confused or kind of beat around the bush. This is the first time I've seen AI say, "No, I I won't do what you're asking. I won't mimic that person's voice. I won't write a fake quote from, you know, Morgan Freeman." And I I highly doubt that 4.7 would

24:02agree to use this skill. Um I do think it's so useful though for examples like what you and I gave. I'm going to try it. For what it's worth, Andrew, I'm going to try it cuz I want a button on my desk that can turn on my air conditioner before I walk out the building. So. I want to be able to use Siri to say, "Turn on the air conditioner." I'm living in Texas. Just

24:20like cool it off before I get there. Again, download the report. It's going to be in the description, and if you have problems like some of you did last week, let us know, and we'll fix it for you. Uh here is free Claude Code. With a name like that, how is it not going to get a

24:34ton of attention? my subscription right now? Like could I just like cancel [laughter] the card and everything? You know what? I [snorts] actually I thought maybe this was the answer, but uh I had a conversation with you. Here's Here's the guy I think who helped get it a lot of attention. Goodbye Claude Code, he writes on um Twitter. Goodbye subscription fees. Somebody just built a proxy that runs Claude Code completely free, and it's wild. You literally just plug in Nvidia's API key and point Claude Code at the local host. That's it. Um tell me what this is. Um and we've obviously talked about it, but let's queue it up for Yeah, so when

25:08you're using Claude Code, it's it's two things. One, it's the command line interface or it's that extension in VS Code. So, that's the user interface, right? Like where you're clicking, you know, what where are you typing. And then the other thing that it is is a machine underneath that has a bunch of tools that knows how to search the web. It's system prompts and and some software. It's not complicated, but some

25:32software that ties all that together. And that's kind of this harness concept we're hearing more and more about. So, it's those together, and then it has to talk to a model, right? Like there's no AI built into Claude Code the the agent. It says, "Well, now I'm going to talk to Opus or Sonnet or whatever." And unfortunately with Claude Code, if you really, really love their interface, their CLI, their extension, and you want to use a Codex model or you want to use a free model or Gemini, you can't. The only models you can select from are the Anthropic models. And so, it's not open like other ones are, like

26:09this one is showing you to be. So, this is like kind of like reverse engineering where they're saying, "Hey, we can figure out how Claude Code is making their API calls. We can get in between that, and now we can send those API calls to Nvidia or we can send them to Qwen or we can send them to whatever

26:26model you want." I want to So, here's here's what I look at when I when This is what you showed me to build in. This is Claude Code in VS Code. I chat here on the right with Claude Code. I have the files that are built for me on the left, and then if I click on one of the files, I get to see what's inside it. This is what you're saying in here, I can now switch the model, but I can only switch within these models. I don't know why Opus 4.7 isn't in here right now, but

26:52okay, I can only switch within these. And you're saying now they're enabling me to switch, but the reason that you say this is kind of useless is because you told me I can go here. I can type in Open Code. You see that I did it. And I essentially install Open Code right here. Trust the publisher and install. And now I will have another conversation right here. This chat right here, I guess, is what it is. Oops, it seems like it's working out of that. I mean, oh, I think you need to have Open Code installed on your computer as well. Just like Claude. Like you would when you

27:23installed the Claude extension, you installed Claude on through the regular terminal or, you know, downloading it, and then you install the extension here. But yes, it is a drop-in replacement. Open Code has been around for a while now. Uh they uh obviously have tons of stars, very high rating, and and I, you know, Open Code already does this. Open Code lets you pick between any model you want. You can run a local model. You can run any

27:48of the API models that are out there. [snorts] There's no restrict. So, should I should I be using this? I mean, honestly, a lot of what I do in Claude Code doesn't need the super sharp brains of of Opus 4.6 or Opus 4.7. I use it because what the hell, it's it's 200 bucks a month, 100 bucks a month. It's worth it for me. But should I just be saving money and switching to simpler codes that are free

28:11or cheap? Yeah, I mean, you could try it. Honestly, like I look at the $200 a month I'm paying for Claude Code and what I'm paying for Codex, and I say, "Hey, like that's worth a lot actually for my time." And I I, you know, I just want the best model. You know, when I'm doing most of my work, I want to work with the smartest person, and I want them to work as fast as possible. And

28:31so, yeah, could I save a couple bucks? Sure. But if it costs me even an extra 30 minutes a month, it's not worth it. The there's so much subsidizing happening in these other models that if you're not taking advantage of the like Claude Code max subscription, um which which by the way, it's $200 a month, and it gives you about $5,000 worth of API credit. That's how much it's subsidized. You you got to take

28:55advantage of that. Um I like the community notes on this. This proxy lets you use the Claude Code interface with alternative AI models via free tiers like Nvidia's NIM, and does not run Anthropic's, which I think he's hyping it up, but I I think that that's the point, and I don't think that this

29:12this repo's hiding it. Fair point. You're saying it could help, but honestly, why do you need Claude Code in order to do that? Why not just go directly with with um what was it called? Or hey, the the Codex setup from OpenAI, they'll let you pick whatever model you want, too. It's kind of just an Anthropic problem. So, if you really, really, really love Anthropic's VS Code extension, and don't want to use any other one, then okay, this makes sense to you. This is your way to to branch out. But there's already lots of other ways to branch out that that don't

29:44require this kind of hack. All right, let's spend just a couple of minutes on this one. Open AI agents, um what is this? Yeah, honestly, Andrew, this is probably worth a full video because what this does is it lets you take the power of Codex or the power of open code or the power of Claude code and it lets you build that power into your apps. So, when I'm building a new AI-powered app these days, I'm not calling LLMs directly. I'm usually using the Claude agent SDK or this like Open AI agent SDK and it gives you all of this like multi-turn capability. It can scrape the

30:23web, all the tools that are built into that. It can run commands locally. It can read and write files. I don't have to write any code for that anymore. And my agent automatically has access to all those tools. And hey, if you've got a bunch of MCP tools, you can connect your custom app to those MCP tools as well. So, anything we're building, we're

30:43building on an SDK like this now. It It doesn't get enough attention. It needs more attention. If you're building an an AI-powered app, you got to be building with something like this these days. If you want more on this, we can spend another hour. Let us know in the comments. All right. Uh this is uh

31:05a Volper? By the way, last week I totally screwed up just saying the word meme. I said meme and I go, "Guys, can we edit that out?" So, now I'm overly sensitive about my pronunciation. Okay, here's here's the one sentence. Shared package manager for coding agent skills. When one agent solves a problem, every other agent on the network inherits the solution. My big issue with this before you explain like how it's useful is they love this whole like gene and bio biology uh framing for it. It bundles genes that ship together. It's It makes it really difficult to follow their uh their GitHub readme files and

31:43understand what it is. Look at this. But what is it? Yeah, I mean, I think you and I have run into this problem a lot where, you know, we work on teams, we work with other people. And let's say I make a skill that helps uh research leads for our business. Or like a real thing is we

32:00have a skill that has our firm's thesis. Our investment thesis is built into it. Well, awesome. I built that. How do I share it with the other people on my team? Do I send it to them as a Slack message, which is mostly what we do? Do we put it in a GitHub repository somehow

32:15and link? But like that needs a model. How do we organize these? His version versus my version. A lot of tools are coming up these days with very different mental maps around how do we organize these shared skills, these shared prompts so that teams can share them, they can build off of them, they can modify them. And like I agree, I'm a little confused. I'm not a biologist, so I don't quite understand the way that these things relate together. But I love the creativity and and how many different people are trying to solve this problem. There's going to be evolution here that will that will create some awesome

32:51solutions. I don't I don't quite understand this one, but uh one just like this is is going to be awesome in the next couple months. All right. Finally, dive into LLMs. This one was much more popular than than I'm making it out to be. It wasn't the 10th most popular. It was really popular this week. Um but it's all in Chinese and I thought about leaving it out. And then I said, "You know what? Um I think it's worth talking about because what we're looking at here is the Chinese teaching themselves and making available to other people how

33:27to learn all these different AI tools. And it's just improving the talent pipeline and that's what we're seeing here. It's It's a country that's really getting deep into this into AI. And that's my takeaway from this. Yeah, I mean, I think you nailed it. Like when I look and see some of the most impressive research and impressive open weight models, it's all coming out of out of China. Like very clearly, this is a a huge investment. There's like a a big

33:57culture of it there. Um I'm a little bit jealous, to be honest. I know, right? To be surrounded by it. But we are here online surrounded by people who are learning this all the time. We've got this whole report including some of the themes that we see uh developing in the comments. Ask for them. We're going to be doing a better a

34:16better job of responding to comments. So, just say hi if you've seen this far so I know that you've seen it this far. And let us know what else you're looking for. Adam, I've actually sent emails personally from my Gmail account to people who who watch this. And I think most of them thought that I was an AI, so they didn't respond. So, I'm not both quickly put your hand in front of your face. This is a good way to prove that it's not an AI video. Look,

34:38[laughter] a real person. Well, this they believe. What they don't believe is that I'm actually sitting there and typing this out. I'm using audio dictation, but still, I want to hear from you. Let us know what's helpful. Let us know what else you're looking for. And we'll see you next week. And bye, everyone.

34:52Bye.