Claude Code Sub-Agents vs ChatGPT Agents vs Manus AI: Live Test

So, I just wanted to try it again because it is very fun. But, uh, like I said, the technical part, it's been a while. But, I think this should be good. I hope so. So, I just wanted to look at um Claude code sub agent today. I think they just did an update where we can set this up very easy. So, that's what one thing we're going to do today. I also want to check out uh

Claude code agents. Oh, chat GPT agents because I saw I have access here to agent mode. This is Norwegian, but basically it's agent mode and I want to compare it to manus or something like that. But uh the first thing I want to look at is kind of cloud code with sub agents. I'm just going to give it like a few minutes to see if everything is working. So, yeah, I can see it looks like the

stream is working. So, that's good. I'm just going to test the chat here. Let me see. I'm just going to have to. Yeah, if anyone can test the chat, that would be great. Let me see. Yeah, nice. Hello, Marma. Looks like it's working. That's good. try to buy something which agent. Yeah, maybe. Hello Eddie Kong v Red Frederick. Nice to meet you. Yeah, custom agents looks very cool. So, let's give it a few minutes and let's

build like a simple sub agent. Uh we're not going to do anything complicated, but I have some just some simple ideas uh I want to do, but I think you can do some serious uh flows here if you wanted to. I had a look on like X and Reddit and people are already creating very advanced uh setups here. So, it looks pretty cool. Uh, but for this video, I just want to kind of make it easy

to understand because I thought I could edit it into like a video. Uh, hello tritos tritos. Ahmed, how uh the cloud API connection status? It went bonkers for the past couple of days. Uh, desktop and code. Yeah, I was trying out was it a few days ago I was going to use cloud code and I just kept getting um I just kept getting overloaded. I think they must have some kind of um some kind of hardware

issues on Antropic. They're probably running low on GPU or maybe not. I'm not sure. Hello Marcus. just added some agents. This uh amazing. Yeah, I think it's pretty cool. So, I was testing this morning. I created like a couple of workflows with the agents and it looks very good. So, I tested kind of I tested one coherent setup that kind of the first workflow is dependent on or the second workflow is dependent on the first one.

Uh and I did some uh parallel agents that also works. Uh John, what is not working? And hello. I can't really read your uh I don't understand. I don't know if that is Chinese, but hello John. What is not working? The agents did work this morning for me at least. So what I found pretty cool is that um this is pretty cool. So context pres preservation. So each sub agents operate in its own context. Hello do

web 3 nice to meet you. I think this is very cool. So this kind of prevent the pollution of the main conversation. So you can kind of separate. So basically each agent has their own system prompt and they have their own context. I think that's pretty cool. And yeah, specialized expertise uh reusability. I create we're going to create some slash commands just to start the agents too. I think that's pretty cool. I did that. Uh each

sub agents can have different tool access levels. That is also very interesting. Uh, I'm not going to use every single feature in this video, but you can use that if you wanted to. So, yeah, I think it's been five minutes now and let's just create a simple agent to begin with here. So, yeah, Cloud Code released the sub agents this morning or I guess they had sub agents for a while, but at least they introduced a

new way of uh using them. So basically we can have context preservation and stuff. Can you also set up uh claw to know itself when to use certain agents instead of having to tell it to use specific agents? I guess you could do that. It should be context aware of what agents it can access. And since each agent has their own instructions, that should be possible. I think I haven't tried that. But basically, I think they

added this part. So the slash agent command, I think that's new. So that's what I wanted to do today. Just build uh some simple agents. Uh, have you tested to do sub agents uh respect the same hooks for the main agent? The context handoff is interesting to hear about. No, I haven't had time to do that. But isn't uh I thought they shared context, didn't they? Let's find out. I think they share context. Or maybe not.

Let's find out. I haven't really read too much about this. If you know, let me know. Do they share context? I guess they don't share uh if you look here, they don't share this context, but uh I guess the main conversation has access to the context. So, we can kind of move it along. We're going to test that with the concurrent workflow. Previously, they had their own context. Yes. Okay. Yeah. Uh but let's just start here

now. Let's build a simple agent. So we're just going to follow this. And I wanted to create uh just a simple agent. I have created a couple. I'm going to go through afterwards. But let's just start here now. So I think the new part is this part. So they made it very easy to build this, right? So, we can just do create a new agent. I'm just going to select on this project, but I guess you

could do personal too, but I'm going to do this project and I'm going to do generate with claude. Uh, so we're just going to do this. But first, let me show you. I have some MCP servers here I want to use in my agents. So I think the first one I wanted to do was uh create uh with the gro server for the exposts just a simple one. So let's do agents create a new agent project

just to keep it simple. I'm just going to do you are an expert agent. Uh agent to find posts on X about cloud code agents. Hello Sunni welcome. Thanks for joining as a member. That's awesome. Thank you. So you are an expert agent to find post on X about cloud code agents. Uh use the Grock server uh MCP. Uh we're just going to keep it very simple. Return the URLs of the X posts you find. Something

like this. Just keep it simple now. So now you can see uh claude is going to generate the agent configuration and I have stored that uh let me zoom in this under agents here claude agents and here is kind of where my agents instruction is getting stored. So we're going to take a look at what uh claude generates here and I thought this was pretty interesting too. We can select tools the agent has access to. So

we can give it uh different kind of tools. We can check we cannot give it access to MCPS. Uh but and we can also do show individual tools. I have a lot of tools here. So I can just uncheck what tools I don't want it to have access to. I think that could be very helpful. Um, but I'm just going to give it access to all tools for now, I think. Doesn't matter too much. All tools.

Yeah. Color. Let's give it a green color. Grock server MCP. Uh, yeah, that is my local MCP server that uh can fetch and find XOSTs. just for context. So this is the agent we're going to create now. So you can see it's called expost finder. We have the description here. The system prompt um description is very long. Is it agent has access to all tools? So I'm just going to save this. Okay. So now you can

see we have the expost finder agent. That's good. So I guess we can just go back. But I wanted to check it out here because now we have this markdown file here that kind of describes describes the agent. So this is pretty comprehensive. We have the system prompt. We have all the MCP tools. That's good. Query optimization output formats. So, by letting Claude create this for us, I think it kind of saves us some time here.

You could, of course, do this manually, but I will. This looks pretty good, right? Let's try it out. So, it should be we could of course create a slash command to start this agent, but I'm just going to go uh use the X postfinder agent. Okay. So, this should trigger I think the agent. Yeah. You can see here expost finder. Yeah, that looks good. So, now we're using the Gro Server MCP and it's searching for what

is it searching for? Trending today. H. Let's see. Let's stop it. Use the finder to find claw code agent posts. Let's try that. Maybe I didn't specify that. I'm not sure. Let's see. Let's see if the query now cloud code agents. Okay, that's good. I did instruct it to return the URLs of the exposts. So what is pretty cool I think afterwards I'm going to try to run this in parallel. Cloco experience review. Okay, it's looking

up different handles. McKe Wrigley, Dan Shipper. Yeah, it looks pretty good. So, I wanted to I wanted to check out I think it was further down here. Best practices. Okay. So, it's completed 40,000 tokens. Okay. Several exposed discussing local agent. Okay. Give me some URLs. This This of course is not very Okay, so here we got some URLs. Let's check this out. Okay, this was maybe not exactly what I was looking for, but I guess the

I guess it's working. That's fine. All right. So, what I wanted to try was to create a par to run some parallels. So, let's create a new agent. create a new one and I want to do uh on trading. So agents are a better way to manage MCPS than calling them directly adding [Music] kind of depends I guess I haven't really tested it enough but uh it could be I'm not quite sure better way to manage

MCPS than calling and directly adding you could do parallel I'm going to show you now you can do some parallel workflows uh with MCPS You are a crypto crypto balance checker. This is another MCP I have. Your task Your task is to check the USDT my stable coins balance using the Solana MCP server. uh return the balance. This is the most simple agent you can have. Like it's just going to use one tool and that is

going to be my stable coin or Solana MCP server. And the only job this agent has is to check the stable coin balance. But I wanted to see if uh we can run this in parallel. Hello, Koshik. Nice to meet you. So, again, I'm just going to give this agent all the tools. Doesn't really matter. I maybe should have just given it the Solana tools, but for now, I'm just going to skip. I am Nino and

we're just trying out the new Claw Code agent setup. So you can see we have created some agents here and I'm going to try to run them in parallel now. That was my ID. So what if we say run the ex exposed uh agent with query GPT5 and check and what did I call it cryptobalance crypto Crypto crypto balance agent in parallel. I think it's possible to run these agents in parallel. It should be. Yeah, you

can see now we are running both at the same time. So, one agent is checking the cryptobalance and the other one is looking for GPT5X posts. So, I think that's pretty interesting. So, let's say you have 10 agents in parallel. That is a lot of you can do a lot of tokens in a short amount of time. And since they have kind of their own context, that should be pretty interesting. I'm definitely going to set up

like a workflow that writes code or does something with code in parallel with these agents. I also have um concurrent setup that I'm going to do after this. Okay, so that was done. So, let's see now. Yeah, we got some posts and we got the cryptobalance. Uh, but then can they come back and talk to each other uh to cross find references a point of interest? There should be some limit for paral parallel agents. Yeah, there

there probably is a limit, I guess. Uh, I saw a Reddit post. Someone I think run 48 agents in parallel or something like that. Let me check that. I think it was in claw AI here. Yeah, let me find it after someone runs some something like 48 agents in parallel here. I ended up creating 49 sub agents. I'm not quite sure, but that is for sure interesting. Running for two and a half hours non-stop. 887,000 tokens

per minute. H. So, dimish music, uh, can I come back and talk to each other to find cross reference points of interest? I'm going to have to look into that. I haven't really tested it too much. Uh, but if we look at the concurrent, uh, workflow, they um, they can share the context after the first agent is done, I think. Let's see if this is working. Yeah, it's working. This is an old one. So, I see

one thing we're kind of missing with the the expost finder. This is the guy who broke the connection. Yeah probably. Hello. Log up. 890,000 tokens per minute. You should run out of Shouldn't you run out of I guess it's running on the max, but still I'm not quite sure. It's crazy. So, I'm going to do a video later about my stable coin MCP. It's very cool. I can send u stable coins with the MCP server around.

I can pay for stuff online using stablecoin. I I I created that uh a few days ago. Uh but I wanted to see so our expost finder is in here. So let me just copy this path here and make some change to this agent. We need to run a date bash command to get today's date before searching because we want only fresh x post, right? So let's see if we can do this to just update. Um

okay, so it's reading this. That's good. Yeah, that's good. Get current date. Yeah, that looks good. So, I'm going to run it again and see if we can only get exposed for like the last 24 hours or something. Exposed agent uh only posts in last 24 hours query. Let's do cloud code agents. So let's see if it runs this uh bash command here now. That is quite important. Okay. So, it's doing this date 25th. That's today,

right? From yesterday today. Yeah, that's good. It also searches internet and search information about the crypto based on your profile. Uh maybe it also searches in that and search information about crypto based. Yeah. So the way I have set up um try to pro uh okay yeah we can try that. So my crypto MCP is just uh check balance of my stable con wallet and check uh or send funds. Found 15 expos about clo agents last

24 hours. Let's check it out. Looks pretty good. Custom sub agents now include code. Perfect. Run/ agents to get started. Yeah, that was from yesterday. Pretty good. More details in the docs. So these uh I don't know if you saw my Grock cloud code video. I really like the way of using Grock API to get the relevant context from uh X.com or Twitter. It's pretty good. I mean I have built something similar to yours one. Okay,

that's cool. Did you invest something? Did you buy some coins? So Ahmed uh try to prompt it without explicitly mentioning to run the agent. Yeah, let's try that. So let's clear uh this uh we can even exit and let's go in again. So, we don't have any. And let's say I need to find some expost last 24 hours about cloud code agents. Now, I do have an MCP server, so I don't know if it's going to

pick the agent or the MCP server. I'm not sure. Let's see. Okay, so it is picking the agent. I said I need to find some expost last 24 hours. So, it's going to pick the agent. Running the bush command and searching. Okay, that's good. Yeah, that's working. Was that what you meant, Ahmed? I think so. Yeah, this is working. Uh, it's prioritizing delegating the task. Yeah. So, it is very interesting and I'm looking forward to pe

seeing what people create uh on building different apps you testing and everything with this. I think it's going to be pretty cool. I'm definitely going to try it out. Not today. I need to look a bit more into it before I I'm going to create my own uh agents to do coding, but I'm definitely going to do that. Yeah, that worked pretty good. Uh but I wanted to test out if you look at agents, you can

see I have a trend surfer detector and a viral video producer. These are the agents I created this morning just to test. So these agents uh look on X, Reddit, and the web to try to find trends and it's going to make a video from those trends, but they are co dependent on each other. So, we need to run the trend surfer one first and then so they can't run uh in parallel, but we can try

them out. I think it's pretty cool. So, let's run them. We can chain them. Add a follow-up question to the main agent. See if the sub agents passes the context fully. I think it should. We can you can see it now when we do the it should it should do that. What I wanted to try was um was this uh chaining sub agents for complex workflow. You can chain multiple agents. So first use the code analyzer

agent to find issues then use the optimizer sub agent to fix them. So that is kind of a chaining. So that is what I wanted to try. First use the trend surfer agent to find trending topics. I don't know. Then use the video agent to create a short video something like this. So, we're going to chain them. I haven't tried this, but I think it should work. It should pass the context. So, the first agent is

going to be the trend surfer detector. So, this is just going to check um x.com, Reddit, and internet web search to find trending topics. Are you still using cursor from time to time to have a little comparison or just cloud code? Uh Matias, I just basically now I have claude code like the max subscription and to be honest I just have the free version of cursor at the moment. I found myself using it not so much

anymore. So basically I was using claw code for almost anything and for me kind of the free version of uh cursor was enough I found out but I might go back to pro I'm not sure yet. What do you use? Do you use cursor or Okay. Okay. So, you can see here now we are using the my Reddit MCP to look up subreddits to find some posts. So, we're going to do a lot of research here

to find trending topics. So, this is the first agent, but we are kind of dependent on the context we get from this to create the next part. So, we should be able to pass on the context, I think. I was okay. So, you can do that. Interesting. Okay. So, that was done. So, we 61,000 tokens. We found some several highscoring drama topics. Voy shutdown. I don't know what that is, but it's probably some kind of internet

thing, but let's try it. Okay, we're going to create a temporary directory. I use cursor, but just for the convenience since I just don't have time. Yeah. Okay. Yeah, I see that. Cairo. Yeah, I haven't tried that yet. There's no looking back at IDE agents, at least for me. Uh, remote SSH dev sessions. Cool. Yeah, the value of the max subscription is crazy on cloud code. I'm scared about cursor. How do you think the pricing of

the max is going to go up? I'm expecting like a $500 subscription coming soon. If we're going to run Parallel Agents and we're going to do what did he do was it 900,000 tokens per minute. They have to they have to do a they they're going to do a higher subscription. people just gonna get dependent on cloud code or other and they're just going to put up the price, right? That's what's gonna happen. Okay, so you

can see the agent is now creating the the video. 70 seconds. That was pretty long, but I'm just going to let it run. We can check here what it's creating. Don't say that. Yeah, I would prefer if they drop new features on Mondays because now we will probably not go outside this weekend. Yeah, they should, right? Okay. So now you can see my agent is doing screenshots of X posts. So you can see this is the

first screenshot. This is a V showjo post. I don't know what that is. We do another post here uh from X. So this agent is gathering screenshots from X using um is it playright or I think it's playright or puppeteer. 2 and a half hours 131 million tokens. Yeah, that's pretty crazy. And it's also generating some images. Yeah, I haven't tested this agent too much, so there might be some bugs here with the video agent, but

let's see what we end up with here. I'm not quite sure, but you can see we did pass on the context, right? We used the first agent to create um to find the topics and it passed on this vtuber something. So it is passing on the context I guess. Well, let's take a look at this post code custom slash commands uh like TypeScript check. He created this and utilitating the new sub agents. So, I guess he

just do slashcomand Typescript checks and it's going to run all of these agents. I'm not quite sure what he's doing here. It's just looking he has Yeah, I'm not quite sure. How do you deal with information overload and checking applicable changes? Sounds like a headache. That's pretty interesting. Okay, so we're going to do this just an ffmpeg command. Have you used Gemini CLI? Does it have agents like CL code? Yeah, I tried it. for being free.

I think it's very good and it's open source. So, I think you can I saw someone use uh the new I think they use the new um they probably forked it or something or used uh is it Quen or Kimmy K2 model on kind of Gemini CLI? Um yeah. what do you call them? Uh CLI agent or like cloud code but uh I have just been using cloud code because I just like it more but it's

a pretty good offer on Gemini CLI I think because of 2.5 pro. Okay, we are generating some music here with agent and yeah I don't know how this is going to turn out. If we go server, we can find maybe or is it temporary? Yeah, here it is. Let me listen to this. [Music] You're not going to believe what just happened to V Showjo. Here's what's happening. V Showjo, the massive virtual talent agency worth MI. Okay.

Yeah. Sorry about the reverb. Let's try now. You're not going to believe what just happened to V Showjo. Here's what's happening. V Showjo, the massive virtual talent agency worth millions, just imploded spectacularly. Yeah, I guess it's working. Hello, Benjamin. Gemini embedding model is now generally available in Gemini API. Okay, I haven't had a look at that. Gemini embedding. Haven't tried it. Are you looking to build some vector vector stuff? Okay. So, it's trying to push this

to YouTube now. I have a pipeline that uploads this to YouTube. Yeah. Uh, V Jojo just betrayed its own stars and lost hundred lost 10 million. So, where did that video go? Let me log on to Let me see if I find the video here. Okay. Yeah, I think it's finished. I'm just going to check if uh we uploaded the video or not. I think we did. Just going to check my mobile. Yeah, we uploaded it.

Um, yeah, we might take a look at it afterwards. You can see uh I successfully used both agents too. Let me just What is the agent brewing? It just made a video. So it did found some trending drama v Jojo. I don't know what that is. uh created a simple YouTube short about this including yeah the video is ready uploaded to YouTube but I wanted to see did I remove it. I think I have some clean

up here so I guess we just deleted it. Uh but I can look it up on YouTube. Give me a second here. Uh after this I think we're gonna I want to try the new agent mode on chat GPT and compare it to manners. That was my idea. Does anyone have any Yeah, thank you. Does anyone have any idea what we should test on chat GPT agent? I haven't really thought about it. If anyone has a

good idea, let me know. Uh, I'm just going to find the the video here. Publish. Okay. Share. apply for jobs. Buy on Uber Eat. I I'm uh I'm going to look up the video. I just um it's on a different channel. Okay. Training ML MLS. What do you mean MLS? Machine learning. Oh, yeah. I'm going to I'm going to open the video now. So, I just uh put it here. Let's find this You're not going to

believe what just happened to V Showjo. Here's what? Okay. Machine learning. No, I haven't looked anything at the machine learning models now. Uh, so let's watch this video. It's It's a bit long. Uh, but let's listen to it. Um, I'm going to turn off my mic here so we don't get the reverb. You're not going to believe what just happened to V Showjo. Here's what's happening. V Showjo, the massive virtual talent agency worth millions, just imploded

spectacularly, and their own streamers are exposing everything. First, V Showjo announced they're shutting down operations, leaving dozens of streamers high and dry. Next, former talents like Bay started dropping bombshells about how management, especially CEO Gun, completely betrayed the people who made them rich. Then, Reddit exploded with over 2,800 up votes as fans discovered the ugly truth about how these streamers were being treated behind closed doors. The impact is massive. Showjo. Here's what's happening. V Showjo, the

massive virtual talent agency worth millions, just imploded spectacularly, and their own streamers are exposing everything. First, V Showjo announced they're shutting down operations, leaving dozens of streamers high and dry. Next, former talents like VBE started dropping bombshells about how management, especially CEO Gun, completely betrayed the people who made them rich. Then Reddit exploded with over 2,800 upvotes as fans discovered the ugly truth about how these streamers were being treated behind closed doors. The impact is massive.

We're talking about a $10 million agency. Okay. Yeah, you kind of get the point. I saw we collected few too little. There was not a lot of images here. It's just four images. That's not too That's not uh enough. Uh, how much does the creation of this video costs with all APIs included? Uh, basically nothing. It's like since we we only gather images, it's yeah, basically nothing. Like maybe 50 cents, maybe something like this, 10 cents.

I'm not sure. So basically it's costs nothing to be honest. Make it check cryp crypto portfolios for top 10 wallets in a coin. Make it check crypto portfolios for top 10 wallets in a coin. meaning to see if we can continue in merch or see if exit autonomous mode random questions. Do you mean yan? Do you mean that for uh shutp agents meaning to see if it'll continue in batch or if it exit but only we

can try that. Uh but yeah someone asked about this workflow here. And it's doesn't cost much. It's a bit on since I'm using cloud code now for uh the to get everything. It's not much. The Grock API. I'm not quite sure, but it's very cheap. Maybe 20 cents per video or something. But this was maybe not the best example. But just to sum up, I think this is super interesting sub agents. I'm definitely going to create

more of those. And we kind of saw that they do share context because this agent first went ahead found some trending topics we can create a video from and it going to pass that on to the video producer agent and it used that to create the video. So it is passing along but I guess if we do parallel parallel agents they won't do that I guess. So if we have these concurrent agents they pass along context

I guess but if we do parallel they probably won't. Hello Eldar. Uh it's been a long time since I did that. Very long time. So unfortunately I don't but there's been a lot of development in uh speech to speech local. I can't remember that model now. I did try it. I think it's like a Korean model or something or at least a Korean some people from Korea created that they have a really good speech to speech

model that you can run locally if you have like 16 maybe you need 24 GB of VRAM I'm not quite sure maybe 16 but no I don't have a new video on it much that's a good on. But isn't that a bit I think that needs more VRAM. It's good though. So yeah, Claude sub a pretty cool. definitely going to try it out more, but I wanted to move on to um trying out chat GPT agents

and Manis. I think I have a lot of credits here on Manis, so we can compare them. I have like 32,000 credits or something. agents token use. Did you do that now? Yeah. Uses a lot of tokens. 9 minutes. Yeah, I think that's the one. Cooro. I think so. Cool. Crazy. I'm not quite sure why it uses so much tokens. Before you move on, can you please tell me how uh you set up Trend Surfer YouTube

upload um and the viral video generator. Yeah, that's going to take a lot of time. But basically, uh, we can clear this to make it very short. What you need to do is I just have I have a MCP server. So, I have an 11 Labs MCP server. I have a Google image search server that can get images from Google. I have a music MCP server that generates music, a Reddit server, a replicate server to generate

images. I have a screen screenshot MCP server that uh does the screenshot part and the YouTube MCP server to upload videos. I guess I have one more Grock. What did you ask about? Yeah, and I have Grock to look on X. So, basically, I'm just chaining a lot of MCP servers. Spent a lot of hours in manus lately. Um, closely building my own due to a bug. Yeah, we're going to compare them. Watch a guy today

do a test with the new Open AI stuff. 40 bucks in five minutes. Uh, where do you get all these MCPS? Uh, I just built them myself. So, they are just local everyone. Uh, it's not that hard to build this yourself. You can just look up um cloud code documentation MCP and just find documentation for let's say Grock API and you can get uh cloud code to just build this MCP servers for yourself. Shouldn't be that

hard. Config file. There's really no config file. It's just uh it's just code, but a lot of people have asked for that. So, I might do an update somewhere where I'm going to upload the code for the MCP servers. There's no config file for this at the moment. It's just code. I don't know why anyone would use anything other than cloud po code at this point. Yeah, I'm kind of in that camp. I use cloud code

for everything now. Just not not just writing code. How does this tie with the sub agents? It doesn't really tie with the sub agents. You can just you can just set the agents to to use the MCP servers you have right on the instructions. I also have some um so I have some slash commands here that starts the agent you can create that too but basically just create your own local MCP servers. Gemini is better. Do

you have the 100 or 200 subscription? I just have the 100 now. I just wanted to start with that just to see just to see uh if I need to go up or down or it's fine with 100. Yeah. What I think is good about Gemini 2.5 Pro is of course the context window. So, if Claude could get the 1 million context window, I think it's going to be pretty good. Can you make a local MCP

server example now? Uh, maybe. Let's see afterwards. I want to try the chat GPT agent. Is a sub agent calling another agent? I guess you could do that but like I said I guess I guess we don't see it here but I just you can do you can do chain calling of agents. So you can just say uh chaining sub agents first do this then do that. Have you tried augment code? No, I haven't. But uh

I guess they they asked me if I wanted to try, but I haven't done that yet. 1 million contacts with Claude solve all my problems. Yeah, I really want that, too. I don't know why SHP Claude and Grock already have moved on to not have moved on to 1 million token context. Yeah, I'm not quite sure. It might be hardware GPU. I think they already are kind of struggling overload sometimes because sometimes in cloud call I

get like overloaded. Augment is good but you can't select LLMs you want. Okay, that's interesting. I guess I wanted to try a new CLI AI agent that can try the new like Kim K2 and Quinn by using open router or something. I think that could be interesting. Open AI agent SDK versus pantic AI. I don't I have so little experience experience with those. I wanted to look into kind of OpenAI agent SDK, but it's just so

much easier just to use clo code or something with the MCP setup. I found that Gemini 2.5 Pro starts to degrade past 100,000 tokens. Yeah, that might be true. But uh my experience has been pretty good with uh big context accuracy dropping up to 200,000. Yeah, that might be true. Are you a software de? Nope. I have like electro electrical engineering background. You can change cloud code agent API to K. Okay, you can guess you can

that. I didn't really think of that. I might So you can use uh the open router API. Have you tried that? V Rick Quen 3 thinking one of the most advanced reasoning models. Yeah, I I checked I wanted to try out the Quen tree, but my issue is sometimes it's really hard to kind of tell the difference right? Gemini 2 had 2 million contacts. Yeah, they stopped that. I remember that. But I was pretty impressed by

the 2.5 context window to be honest when I have tried it. Pantic AI with small local LM. That's pretty cool. What's your advice for content creator in the AI agent space? Um, advice for content creat uh, Hussein or do you mean like X or Okay cool. I would say try to catch some trends. If you see like something new coming up, just that's a easy way to get some views and maybe you get someone to stick

around. Catch some trends maybe if you're struggling with getting people to watch. I think that's a good advice at the beginning and then try to create your own style. I guess Kim K2 is the best free tool. uh free tool calling out there. Yeah, I heard good things about it. N8. Yeah, I heard a lot about that. I've just tried it briefly, but have anyone of you tried N8? I think it's Can you create any mode

with that? Um I'm not quite sure. It looks interesting though. But isn't it quite quite slow? Everything is is can you do parallel with NAN async? Okay, but we need to move on, right? So, let's uh a lot of nice questions though. But while we N has many models to choose from. Yeah. text to speech in Norwegian. Uh I just use English you are but uh text to speech in Norwegian. Uh 11 Labs maybe Google too

Kimmy to be specialized for coding to reduce its size. Yeah. Are there any good uh quantized version RV of Kimmy? NAN is not bad for rapid prototyping. Okay. Yeah, I've been just using MCPS instead of N8 on code. You can have an MCP that gets no data. Okay. I might have to I've been postponing N8N for like eight months now. I think best I got is two bit 300 gigabytes. Okay. Yeah, it's going to be rough

to run that locally. I guess I don't know how much you you lose if you go two bit. But open router this should be still cheap. Let's check it. Uh, open router. Do they have the price here on uh kind of burns my eyes here? Um models. Do they have Kim K2 free? Okay. Is that free for now? What GPU do you have? I have a uh not on my laptop. This is the Mac, but on

my desktop I have um Nvidia 4080. I have deployed N8 using Docker. It's super simple and you can host locally. Okay. Yeah, I might have to take a look at it. A customer asked. So, I tried to build a football soccer prediction with LLMs to get a feel. Okay. Sapier. Yeah, I've tried Sapier. It's They have an MCP option, don't they? I think so. So, on open router, it says Kim K23. H really. Okay. So, here's

the price. Yeah, they have a but they also have a paid one. 14 cents per million in, 250 out. The free version is pretty bad on open order. Yeah, I was thinking that since they have a another option. Let's check out Quen. They have a free option for Quen 32. This is the paid one, I guess. So that is.3 3 cents in 3 cents out. Okay. But what I heard about Kim K2 is the tool use

is very good or instruction or tool calling. All right. That is interesting. 30 30 cents. Yeah. Sorry. But I heard that tool calling on Kim K2 is very good and that is very important on cloud code, right? Tool calling cloud code has by far the the cloud models is by far the best uh tool calling models. Maybe K2 is good too. Yeah, Kim K2 fine-tuned for agentic use. I think that's pretty interesting. makes it good for

this uh agentic workflows. I'm building a trading bot with ML machine learning and clo code today. Getting disconnected from IB gateway. H not quite sure. Maybe some other guys in the chat have an ID. I haven't tried that before. But uh we need to start something on CL chat GPD agent here. So what should we do here? Let's do some few ideas around what we should do with uh claw chachi the agent and let's do the

same prompt on manners so we can compare them. So if you have any idea what we should do on chat GPT agent and compare it on manus let me know. We have some suggestions here. Find a vintage Japanese style lamp. Oh, that's boring. Find a Lego set under $60. Create a spreadsheet. I just tested the prompt. Find top 10 wallets for the uh Ethereum cryptocoin lister holdings in a table. Okay, we can try that. That's a

simple one, right? Find top 10 wallets uh for the Ethereum uh cryptocoin uh list their holdings in a table. We can do that. So, let's do that on agent and let's do it on manus. Okay. Uh, we can do quality start here. Start. I haven't tried agent before. I don't think you need an agent for that. No, but let's just see what happens. We can try other stuff, too. I'm just going to update the title. My

keyboard is disconnected. Chat GPT agent versus manus AI. Okay. I think they're um Oh, okay. So, you can see Manus is looking at nansen.ai. I like their UI here. Looks pretty good. It's in Norwegian. Sorry about that. But uh it's going to coin edition. Manus is already looking up. Manus has already found the top 10 wallets. I guess that was pretty quick. It's still looking. So, Yan, can you confirm this? Does this look correct? It's done.

Okay. So, they were quite similar. They were pretty similar here. Bitfinex is number 10. Yeah. Okay. So, that was pretty similar. Not a big difference there. Uh, let's do something that's a bit more challenging. Same spots. Yeah, looks pretty good. So, let's do a new task. Took about a couple of minutes. Not too bad. uh agent mode. I'm going to try to come up with um I want to do let's do let's do some research on

let's do something from today. So let's do some something on let's do something on cloud code agents just for fun. Um find a new feature uh slash agent uh on cloth code. There's been some questions whether Chhatship the agents add much value or just using deep research alone. Yeah. Uh my plan was not to use uh I have no intention of using agent uh mode on Chipity. I'm just going to try it now. Find a new

feature agents on cloud code. Create a stepbystep step guide. turn it into a can you even do that website? I don't know. Create some slides. I think I find it so boring. Uh does anyone else have any idea what we can compare manus and shat the agents on? It's so yeah, find a new feature. Create a let's find some X post then I guess to I guess it can't access Reddit, can it? So let's think. So

it has access to deep research and it can do computer vision. Do some research. Yeah, I think we need to do some research because the the idea is that it can Can you schedule the agent? I don't think so. Haven't thought about that. I don't think so. So, it can do research uh based on deep research and it can do computer vision. Okay. So, I think we need I think we need to do some plotting. Find

the changes in Bitcoin. Make a video. Can it make videos? Chat GB agent. I don't think so. Uh, find the changes in Bitcoin price last 24 hours. Build a uh plot changes into a This is boring spreadsheet. create some visualization of the changes graph uh presented in plotting is easier on noise format I don't know just let's just try that I find it so boring. Uh, but Manis, uh, I did a video on it, like a

sponsored video not so long ago. Was pretty good. I created like a nice VO3 video using Manas. Was pretty cool. I'll do the following. Find and access BTC price. Collect spreadsheet visualization findings. Deliver. Okay. and agent is going to I will research the price up to date resources gather data compile into a spreadsheet create a graph and present it pretty much the same so that's good okay so it's creating some code here using the API while

Manis is uh searching for an API Bitcoin is also writing some code ship will just look like fluff. I think they did a pretty good job on the UI here, doesn't they? It looks pretty cool. Okay, so you can see we are creating a Python file in Manas. I like their UI too. And we can go back here and check, right? Jump to live. I think it's the best UI I've seen for operator like a chance.

Yeah, it looks pretty good. Think did a good job here. Okay, so we ran the code here. We found 24 low, 24 high. It's interesting. Manis is already doing work. Yeah, I like Manis. I think it's pretty good. I don't know what's going on here. It's just reading. It's just read. I don't know what it's reading. This looks much more on point. Manus gives me better results versus GP agent. Yeah, I think so too. It's just

writing a lot of code. It's going to probably going to plot. Decorating the spreadsheet. Why can I scroll up here? It's a bit strange. The Bitcoin data script successfully created 127 hourly data points from Yahoo Finance. Okay. Uh, SH GPT is still struggling a bit, I think, but it's kind of hard to tell. Doesn't really doesn't really say how far it's come. Here we can kind of see what we have done so far. I kind of

like that. I don't know how well you can see this but we are using is it seaborn loading the bitcoin data it's still reading here's a good one which require multiple agents give a location on Google map and ask which business will be the best for startup up. Yeah, maybe we can do that. Manus is really good because it works with files to transport state and therefore does not overload the context as much. The research is

also pretty solid. Yeah, I think so. This site can be reached. Okay, now we got something here. We're looking at blockchain. I don't know what's going on here. Okay. So, we ran some code here. We loaded 37 data points. Okay. Now, we're going to do some slides. It looks like Manus is much quicker at least. Well, I'm not sure. Uh, I'll be right back. I'm just going to But you can't edit messages to go back like

in cursor and if the container breaks, the chat is simply gone. Okay, I haven't really So, can I do anything here in chat? follow up. It's looking at some uh logos, creative problem solving with AI. Yeah, I'm kind of I'm kind of bullish on computer use, but it has to be background computer use because it's very slow, right? Okay, so I I'll be right back. Just going to grab something Okay. So something is happening. What about

Okay. So ch man is making a video. It is. It's making slides though. Looks pretty good. I'm looking forward to seeing the result here. Let me check my camera. This looks awesome, right? How is SH GPT doing? It's doing something. It's at least creating some kind of code here now. Or is it? Yeah, I think it is slides. Manus has completed the current task. Okay, so let's check it out. Okay, so this is the first slide.

Bitcoin price analyst. This is from today. We have I just had the open AI and try to pro the script based on how okay so it's watching YouTube videos. Let me see. Okay. The fi really so both finished. It took eight minutes but look at this. Look at the difference. So, let's check out Manus now. So, this was the first slide. That's pretty cool. You got the logo, 24-hour market performance, overview, and key findings. So, we

can see kind of the graph here and the total change. That looks pretty good, right? I like the style here. We got the 24 high, 24 low, the range, volatility. We got this uh here we got the plot thing that looks much better I think than the we have some key observations market implications. We have a candlestick I guess. Is this volume? I'm not quite sure. Okay, so we have volume too. That's pretty good. So we

can see there's a lot of buying at the end here. Price change analysis. statistical analysis. It really went hard in on this, right? This was just for the last 24 hours comprehensive dashboard. It's repeating a lot. But if we compare it to this, it looks think it's a I'm not quite sure if did the price go all the way down to 116 or 114? It might have done. I'm not quite sure. Yeah, the current price. Okay.

So, this looks quite boring. What do you think? Who won this? Has to be Manis, right? Not even close. Kind of impressed to be honest. Did a really good job. Pretty cool. We got the the spreadsheet here. Yeah, I'm kind of sold on manners. Yeah, I think it's pretty cool. Let's see how many credits we spent. We had like 37,000, I think. 32,000. I don't know how much we spent. view usage. We spent 180. That was

not much. Yeah. So, I definitely would give this to Manus. And we can also do on Manus, we can do video. We can create V3 videos. But I guess I guess it did get the correct numbers. But I don't think I would use this in my slide deck. Too boring. I don't know. kind of sure I would pick manis any day. Interesting. Uh let's see. Let's open up the chat again here. Okay. Yeah. If you have

any anything else you want to test on manus or on chat GPT, I guess we can try the video. It is pretty cool. So, what I did in my my manus video a few weeks ago, I first did some research and I turn it into like a video. That was pretty cool. Um, let's see if we can think we can continue on this now with a video. But if we can we download this spreadsheet? Yeah, we

can. Let's do a new task and let's do video and let's upload the spreadsheet, right? Viral video about the Bitcoin. Yeah, that was my idea. present the BTC uh price changes in uh 8sec video. It's a bit short though. 8 seconds uh over two 8sec video clips. VO3. I don't know if this is going to work, but let's try it out. So, we uploaded the spreadsheet with the numbers, and let's see if we can turn this

into a video. So, it's going to analyze the Bitcoin data. create a visualization. Generate the clips. Yeah. Okay. I'm not quite sure how it's going to do that, but let's see. Okay. So, it's analyzing the data first. Yeah. I'm not going to even I'm not even going to try this on because we can't generate videos on chat t. It has to kind of turn this into some kind of prompt or script. I guess Yeah, it didn't

really have to do this. Why is it doing all of this? That was not That was not the point. Why is it creating Okay, we're just going to let it cook here. And I wanted to see if anyone else has talked about talked about uh survey chance. Yeah, kind of looks like this. Pretty cool. Okay. So, generate your video using quality mode. Um, yeah. So, you can see this is pretty expensive. 600 credits for 8 seconds.

But we're going to use quality mode. That is V3. So, it is quite expensive. But how is this going to handle turning the Bitcoin data points into a video? This might be a disaster, right? Yeah, this is probably going to take some time. I think V3 fast is very close to quality. Yeah, I use V3 fast all the time. We can do some videos afterwards. I have um flow Open AI agent will actually try to create

a video voice over. Okay. So what uh does it use the open AI I guess API or something? We can do that. So we're going to select agent mode. create a voice over script from this. Let's see what happens. I can log into flow here. We can do some videos. So, I use VO3 fast. It's only 20 credits, right? It's pretty good. Still working. Let's see what happens here. Okay. So, it's going to create a voiceover

script. Did I say that? Okay. So, I'm going to stop that. voice over script plus voice dot uh mp3 but yeah I don't know if you saw my video yeah okay so it's thinking now I did a video a couple of days ago where I kind of recreated my previous viral Tik Tok video and I said I don't think it's going to go viral this time. I just did basically the exact same video two times. Not

exact same video, but exact same cont concept and it's almost at 100,000 views. I don't know. That's pretty strange. You create a viral video first, then I just copy exactly what I did and I still get a lot of almost 100,000 views now in one day. That's pretty strange, right? Okay, still generating. So now it's going to do professional sounding voice reading the script. Okay. But how is it going to do that? What is your Tik

Tok about? It's about nothing. I'm just testing AI videos for my I have this uh uh I have this course platform here, right? So to kind of learn what works and stuff, uh I just need to create AI videos to see what get views. So this Tik Tok channel is just testing out different what goes viral, right? So I can update the course to kind of Yeah. tell people what works. Viral Tik Toks just some dude

playing with AI. Yeah. But what you can do when you find you can clearly see this concept is working right. this type of video is working. You can see we got 80,000 and 900,000 views. So, we can kind of take this concept into promoting our product or software or anything. That is kind of the idea, right? This took a lot of time. Five minutes and still no video. How much money do you earn with 900,000 clicks

on short? Uh, I don't earn any money. I don't I'm not monetized, so I don't know. Yeah, you can just go to what is the channel? Here, just go to this channel here if you want. Real secret findings. I just took like an old channel or something if you want to see the video. There's copyrighted music, so I can't play it here. Okay, so Manus has completed the video. I don't think this is going to be

good. What the hell? Why is it so? Why is it so big? I can't get out of this. What is going on? So, I'm going to have to download this. Okay, here we can see it. That's not good. Do you think that was cool? Let's watch the other one. [Music] Now we need we need a female news female to present the numbers one video first one. Yeah, that was better. But you can't use V3 for Okay,

so we're going to do a female presenter on the data now. Let's see if that's turns out better. Uh, we can use female voice. That's fine. How is looking for? It's looking for free TTS online. EdgeTS text to speech a hugging face. So, it's trying to find a free TTS. Testing TTS API access. Okay, that's a way to do it. But why is it creating a 7 ft tall? Yeah. Yeah, we should have done that. How

to reduce errors and bugs while coding with claw code? Oh, that's a big question. Uh, I guess. Okay, so we're going to generate this clip. How to reduce errors and bugs while coding with cloud code. I think it has a lot to do with kind of the the framework you have kind of let's say you wanted to create some nextjs project or something like that. I would start by gathering relevant context. Right? So you have like

the upto-date documentation. Yeah, context 7 that's good. So if you have the latest documentation in context working in small chunks, that's a good one too. Uh I like to use modular approach. So you kind of build your project in modules, not in like one single mono file. I think that's pretty good. But finding bugs and I don't know testing I guess there's so much stuff you can do I guess so it's kind of overwhelming to what

is the best way to do it but I would say up to-ate documentation in a context modular approach a lot of people do this t testdriven development I think that's pretty good So testdriven uh development. So you can see this is kind of approach where tests are written before actual code. I think a lot of people use that T d TDD. But since I'm I'm not running any big project. Yeah. You can test it on a

per function basis. But since I'm not running any big projects or SAS or something like that, I don't really I only have small project projects. So chat GPT is still working to find a free TTS. Uh I'm building a web app with cloud code and so far don't know anything coding. Do you think it's possible to do that and do you know coding it? There was a paper that the only killer cover that shows us understanding

that understanding is reduced significantly when you stuff context with too much data. Yeah, I guess so. Uh but to your question possible um yeah it's possible to be build a web app with cloud code for sure don't know any coding it is possible but maybe you could start on you could use cloud code but if you wanted to get started maybe try lava all if you wanted to have some more help but it is possible in

clo code open AI creating video is not great I don't know what it's doing here it's trying to find out text to speech could I give shot my 11 labs access somehow so it could create a TTS audio is still trying Okay. Um, shy b your 11 labs access. I don't think so. But you can do mcp uh delvatio. If you look up uh mcps, you can connect. I have an mcp 11 labs mcp server here

that can generate TTS. So I can do say hello hello chat with uh 11 labs. Right. So when I have my MCP server here, I can just create this uh play it. >> Hello chat. Is it? >> Hello chat. >> Yeah, John. Claude code plus task matters with context 7. Good specs TTD. Yeah, that seems like a good combo, right? Cloud to plan out the process. Yeah, you if you're on clo you can do plan or

is it it's not plan it's uh shift tab tab then you have plan mode then you won't edit any if you do shift tab tab you can go to plan mode so you don't edit any files while you're making your plan you don't any tutorial how to do the setup the 11 labs MC MCP? No, I don't think so. But um it shouldn't be that hard. If you want to create an MCP, you can go to

It's still trying. Oh my god. Okay, so the video is done here. Hello, I'm here. Hello, I'm here to present the latest Bitcoin price analysis from the P. Hello, I'm here to present the latest. What is wrong with this? Hello, I'm here to present the latest Bitcoin price analysis from the past 24 hours. Bitcoin has experienced a notable decline over the last day. The price started at $118,486.98 and has dropped to $116,569. That was not a

that was not a VO3 video. The more you work with Claude, you will learn how to use it. Yeah. Yeah. Sorry about Sorry about the reverb. Uh it's going from the my speaker to the Yeah, sorry. Yeah, this was not so good. Uh but uh I wanted to talk about one more thing or I wanted to try one more thing and that is um let me show you. I had a look this morning at um doing

some JSON prompts with V3. Okay, so SHP is finally trying to create an audio here. But what I wanted to do was to create some JSON prompts for V3. I wanted to test it out. We can try this one. This is just like a template. So I haven't tried this, but basically it's just another way of prompting V3. a more structured way, but I don't know. Well, let's just send this. Let's see what happens. If anyone

has any V3 videos they want to create, just uh just write in the chat and we can do it. Yeah. I will test the t. How long has chatb been trying to do this now? It must be like 10 minutes or something like this. I'm just going to let it run in the background. Hello son. Yeah. Does anyone have any V3 video? Just give me an ID here so we can test out this JSON prompts here.

We can watch this first. Yeah, it looks pretty good. Shut it be around for 12 minutes. produce garbage video without sound. Manage is better, right? Yeah, I think so. Now it's on a Now it's trying to do a lovevoice.com. It's on this. I don't know what's going to happen here. Yeah, we can do that. Uh, create a underwater shot of a dolphin. Camera follows dolphin as it jumps out of the water. Okay, same JSON prompt style.

So, I'm going to compare doing a JSON prompt style and just natural language. So, let's do this. So, let's just do underwater shot. like this just text and let's try the JSON prompt. Okay. See if there's any difference. It's looking for another AI. Yeah, I knew I know it gave up. It just gave me a a voice over script in text form. It gave me this. No, it didn't generate. To create a voice over audio, I

can use an online texttospech service. Enter a script into a web form. Click. Would you like me to proceed? Yes. What are your experiment experience with JSON prompts? If the JSON soft works well, I guess you can have it first create JSON based on the Yeah, I guess so. That's could be could be good. The agent wants you to do the work. Yeah, it just gave up. But let's see now. So this was the dolphin without

with just natural text. Okay, that was a bit strange. Um, I can see it's much okay. So, let's watch this one. This was the JSON prompt. It's slow. Yeah. Uh, I don't know. I didn't really see any difference. A B3 bit of a western style movie. Yeah, we can try that. I forgot I have a chat on my other computer. Uh but let's try it. So, I didn't really see any difference um on that. create a

prompt of Yeah, the turn from left to right looks strange. Yeah, I agree. A western style movie hero breaks uh open the door. All the bad guys drinking takes out their guns and fire. Shouldn't we have any dialogue? Add fitting dialogue. So, let's try the JSON on that. Let's see how chat GPT is working. It says, "Visiting love.com. Paste the content into a text box." I'm not going to do that. You're the agent. Do it for

me. Okay. So, it's not in agent mode now. Why? I'm just going to give up. Let's see if uh let's see if Manus can do this for me. I'm going to give it the file. Let's see if it can do it. Yeah, it's lazy. It's just giving. What is this? It created an image. I need I need a JSON prompt. JSON video prompt. What? same. Is there no context? Okay, here we go. Okay, that looks better.

Uh, it's a pretty long conversation here, but let's try it. Okay, so you can see Manis is doing this. It's going to love voice, pasting in the script. This is what I wanted ChachiBT to do. It just gave up. It had some issues though. It's trying. Much better. I wouldn't say it's the most efficient way to do this. Yeah probably. Okay. Could it click on generate? Let's see. It's going to select a voice. So, it's going

to do. It's going to do a female voice, right? It selected a guy. Okay. Is he going to click on generate though? Oh, I need to verify. The AI tells you to do it yourself. Yeah. Can it pass this cloud for verify? I don't think so. Clicking I don't know if it did. Let's come back to it. Okay, so this prompt violated is probably because of the guns. Um, let's try again. I think he gave up

here. [Music] Let's try without guns. Be we have to get credit. Yeah, I know. I know. I've created a lot of videos, but it looks like maybe it's fine now. Let's see. It's still trying. I think we're just going to stop this. I don't know if it can verify Cloud Flare. I don't think so. We're in this uh Maybe 10 to 20 videos to get what you want sometimes. Yeah, I know it's uh but the nice

thing when we Okay, so we got this. When we select um when we select uh fast V3, it doesn't it's not that expensive. It's 20 cents for a fast video. Okay. So, let's watch it. >> This ends now. Well, look at what the cat dragged in. >> He ain't going to leave here alive, boys. This ends now. Well, look at what the cat dragged in. He ain't going to leave here alive, boys. This ends now. Well,

look at what the Okay, that was pretty strange. That was funny. This ends now. >> Can sometimes cost three, four bucks to make. Yeah, I know. So, this morning I looked at this JSON prompting thing. I thought maybe this could give me like some but a better way to do this is uh let's say we create create a image of a female in her 20ies doing a vlog selfie in on a Safari or something like that

916. So if we use Google to create an image for us, we can use uh frame to video, right? That's going to save us a lot of hassle. It started good. Yeah, I agree. Uh yeah, it doesn't really matter. Let's just download this. So what we can do, we can take this image, right? Let's go to AI Studio. Nice thing about using JSON is you can generate your own prompts with a simple Python script. Yeah, that's

interesting, I guess. Uh let's just copy this um example JSON prompt. Let's paste in this. Did we download the video image? Yeah. Let's upload the image. All right. turn this image into a JSON prompt of a vlogger on a Safari. I don't really Let's see if we can do that. And then we can do we can upload this as a frame. So, I haven't really tested enough if um yeah, I can update the title. Uh V

out just going to update the title of the stream. My keyboard is disconnected. See if I can fix it. Okay, so you can see now we get the JSON prompt. We have the frame here. Let's send that. How do I connect my keyboard? Okay. 71 6. So now we can kind of do we can do the JSON prompt and we can do the frame to image. So this is going to be pretty good. So, we're going

to take this image as an input. Did you try stuff like trip booking and finding the cheapest flight in a time slots? No. I hate those uh benchmarks. Booking a air ticket. I think that's the stupidest benchmarks ever. That takes me Why do I need an agent to book a flight for me? I do it maybe once a month. the results with JSON in my case are better. Okay, that's interesting. I did try this morning. Yeah,

you agree, Edgar. This these benchmarks like I saw I think it was OpenAI did that agent video live stream. Uh looking for something to wear at a wedding. Why do you need an agent? You just need a suit, right? You don't need an agent to find that. So strange. Okay, so let's see what we got here now. Wow, you guys. I can't believe I'm actually here. And look behind me. Actual elephants. Wow, you guys. I can't

believe I'm actually here and look behind me. Actual elephants. Okay. Yeah, that worked pretty good. Yeah, it's not bad. I really like the frame to image. We can do a lot of cool stuff with that. It's not too bad. And we can uh Yeah, let's try this one. So, let's download that. Is that a still widescreen video though? No, I guess it's it is still widescreen, but if you use a video editor, you can kind of

get it 916. You just have to do like Premiere Pro or something and just zoom in right? So if we close this do like a new one. We can try that afterwards. Did I download that? Uh let's do use this uh JSON prompt walking. Let's try again. I don't really like these videos when you're just standing still. Okay. So, let's see if we can get I don't uh Mick Bab, did you did you use the JSON

prompt with uh text to video? I guess so. That is what I tried this morning. It looked pretty good. Changed elephants to dinosaurs. Okay, so let's try this. I want to see if she's going to walk. >> Wow, you guys. I can't believe Let's try to create a a video again using JSON prompts with manners. When using V uh V3, always use uh JSON prompts. Uh X. Let's do an example. do some research on um the

new Claw Code agent feature. Create a video to pre V3 video to talk about it. I don't know. Let's try that. So yeah, uh I don't know if you uh someone here uses cloud code, what um what is kind of your go to MCP servers? on this part. I just have I don't have all my MCP servers here. This is just kind of local to this project. I have my Gmail. I have Gemini, GitHub, Firecrawl, my

Discord, music, Reddit. We can test the Reddit. It's a pretty cool MCP server. Find the Sequential Thinking Context 7. Yeah, I've used both of those. Those are pretty cool. Find the latest posts in latest 10 hot posts in singularity subreddit. I really like this. I created this um Reddit MCP so I can write post, I can leave comments and everything. We can try it out. Well, let's watch this one now. >> Hey everyone, welcome to the

Serengeti. We're on safari and look, a jeep. Hey everyone, welcome to the Serengeti. We're on safari and look, a jeep. >> Hey. >> Yeah, that was pretty good. So, what I want to show you is if we download this now, we can just bring it in here, right? And you will see it's like v. What did I do now? That's the wrong one. I meant God, this one. The hell? Where is it? Yeah whatever. Didn't I

download it? Oh, here it is. So, if you're going to use this as like Tik Tok or something, Instagram, just change this to 1080 1920. Yeah. And just zoom in, right? Something like this. like this. And there you have it. A jeep. Lol. Yeah, that was pretty good. So, yeah, here's kind of my Reddit MCP server. So, here we can find the latest 10 posts from Singularity. We can do uh what about SH GP or open

AI and we can select one of these posts. Let's do claude. Claude AI server. So to create this MCP server, I just there's something called a Reddit app. So you can um connect your Reddits. So here is the post like number nine here. So read post nine. This was the one we looked at that has this sub agents. Yeah, I don't think we can use fire crawl on this. We can't. But here you can see we

get the summary of the post and leave a smart comment. So I can make a comment here using the MCP server. So, we're going to comment 887,000 tokens across 49 subentient is basically running a small data center worth of AI compute. Let's just comment this and let's find it. So, let's refresh this. I don't know if it worked. Yeah, here it is. Yeah. Yeah, it works pretty good. But I'm not leaving the comment. It's It's just

the Claude code that Claude that writes the comments. We can do posts, too. But uh yeah, it's a pretty neat server. If you want to get context from Reddit, I think it's pretty interesting. Did you try some POV video in com combination with audio? POV video in combination with audio. Uh, we can try that. Create a POV image from a jeep looking at lions hands showing. Let's try that. I think I tried it. It's pretty good.

But let's see if we get I forgot to put in vertical here, but let's see. Yeah, maybe that was a bit strange. Let's try 916. Do you want like dialogue? We can try dialogue. It's not exactly POV, but let's try it just for fun. Same with this image. POV fitting dialogue. Okay. Upload um JSON format. Okay. I haven't really tried a lot of POE, but I know it's pretty popular on like YouTube shorts and stuff. It

is interesting. So, our video now has gained 3,000 views since the last time we checked. How is Manus doing? Okay. So, we're going to select quality. Make an online calculator. That is always wrong. Hello Annie. We can do that. So, let's leave CL code. That's the MDR calc wrong uh calc wrong uh claude. Let's create a cal later. That is always wrong. Choose a simple text stack. Consider C++. Let's open up the folder here. Let's try

it. So, we're going to do a main CPP uh implement the wrong arith arithmetic. Uh let's uh create a make file. Uh let's compile it. I'm on the $100. I'm not I'm not on a 200 yet. Okay. So, it's working. 5 + 3 = 9. Uh, create uh C++ UI. Yeah. And I'm on the max, but I'm on the $100 version. 5 + 3 = 9. Okay, so we have our prompt. Let's grab our image. store

prompts. Okay. >> Hey everyone, welcome to the Serenetti. >> God, this took a lot of time. Yes. Yeah, you can say that on it. So, it's pretty interesting. It would be pretty cool to try Quen 3 here on open router. I might do that tomorrow. Could be pretty cool. Okay, so it looks like they compiled. Let's run it. So let's see now. So this is the wrong calculator. So five doesn't look like it's working. Um We

can't input any integers. We walk around thinking l acknowledge the logic of everything. Gathering context. I only leave my house when I hit a usage break or go to work. That's pretty funny. It's going to be very interesting to see if you can have these agents running 24/7. I guess you in theory could now, but like uh it's just going to cost so much. Okay, so it's working. I don't know who wanted this calculator, but uh

so let's do 5 * 5 = 26. Let's do 10 + 10. Is that going to be 21 then? Yeah. So, it's just adding one. Fire would have been discovered in a weekend if the caveman this AI. Yeah, that's pretty good. Yeah. 5. Can we do float? I guess not. Five divided by six. I can do float. Okay. Yeah, this worked pretty good. You want to buy it the code? 0 / 0 = 42. 0 *

0 = 1. That was pretty good. Yeah marked. Let's check our POV video here. >> Okay, nobody move. Just look at him. Incredible. Okay, nobody move. Just look at him. Incredible. That was pretty good. Let's change it up to the lion attacks the jeep. Change prompt. It says generate audio. True. Okay. So, it said whispering. Okay. Nobody moved. Just look at him. Incredible. That was pretty good. the the Greek mathematicians like Purus would have been trolled

with the wrong AI calculators. Yeah. So, let's try this one and see if the lion is going to attack the jeep. suggest new features for our calculator. So let's So one I wanted to do one project one day where we just we create like a very basic app or website and then we just going to let agents um add new features like just add tons of features. Uh, I want an animated picture. So, here's the claw

code research we did. So, let's check out the video. I think I'm just going to do Yeah, I think that's a pretty interesting idea. I'm going to try that soon. Let's watch this. This new technology will change the way we do coding, opening up all kinds of possibilities. It didn't even mention cloud code change view right uh in the jeep is sitting a lion. Okay, we can try that. Happy pan camera. Right. A lion is in

the passenger seat. Yeah, I agree. So, let's see if the lion attacks the jeep now. Whoa whoa >> go. I guess the zoom was a bit strange, but it was pretty cool though. I kind of like I like this look. Where was it? Here. That was pretty good. So, let's see if we take this prompt and let's see if there's a lion in the front seat. So, let's do the same frame. Let's do this. Whoa. Whoa.

But it was pretty good. So, now we're going to do some animations on our calculator here. But what kind of animations? Okay. What was that? So, we got 0 divided by zero. So, we're going to do a lot of animations in here. 1 1* 1. I don't know if I want to call that an animation, but okay. Yeah, that was pretty fun. Can you try to extend some videos? Uh, last frame and running it. Yeah, we

can do that. It's not that hard. I can just I'll show you. So, let's say we took this one, right? Since we already have it here, we just we go here. It's going to turn off the sound. And we just do like this. And I just do this. And I just do this. Right. Save it. Give the character some ads. Okay, we can do that. create some ads for the AI video course at H web the

AI videocourse.com before getting the result. So, what I did here is I just took the last frame on this video. We're going to go back to VO and just upload this last frame. Right. So, we grab this and then we can continue. So let's uh let's copy this. continue with the scene. She turns around and sees a huge rhino, freaks out, right? Let's try that. Okay. So, we're going to fetch. So, let's see if the lion

is in the front seat now. What kind of car was that? Let's try that again. Integrate this into the calculator. What role does Claude play in all of this? You mean in this or in? We did. We already did the claw H and stuff. I'm just hanging out. So if you want to continue, we can just select this next frame, paste in our prompt. Right. Do you mean include code? It's uh natively. It's yeah, it's kind

of natively connected, right? Uh I'm logged into cloud code to through the my plan, but you can also put in your API key. So if you go to models here, I guess we can let this run first. You can see we can make this Let's try this first. I can show you. Five * 5. There was no add here. If you do slashmodels you can see we can select default opus or set. So I'm I'm just

using set for now because I want to save my opus uh if I'm going to do some coding. So I just have set on uh for default and if I'm going to do some coding I'm just going to switch to opus. Hello Alex. Uh I use cursor uh vibe tunnel. I haven't tried that. Uh remote access to cloud code. I tried it. It's pretty cool but I I don't use it since I don't have any big

projects at the moment. I have the free version of cursor but um I don't have the pro version anymore. What do you use? Integrate ads into the calculator. Let's see if we have Yeah, we have this. Now I'm going to watch it. Oh my god. Okay, that looks a bit strange. You can see it's a bit It kind of loses the quality a bit, right? But if we download this now and we kind of go back

to Premiere Pro, we can take this and we can stitch them together, right? What is the biggest project you put together with clothes help? For me, um I think it's just my course platform. So I built this uh course platform here that is hosting all my videos. It's doing Stripe and it's doing uh it's hosting all my course videos. Yeah, Stripe. We have logging. We have databases. Google I think just that's my biggest project. It's on

Versel. And I haven't really done any big projects, but uh this works pretty good. I've sold a lot of this, so so far no errors. Are you available for cons consultancy? We can have a chat if you have anything. That's just send me an email. Uh I can give you an email. I don't know why my keyboard keeps disconnecting. Let me just add my keyboard again. It says it's going to play an add every third calculations.

I don't think so. Doesn't work. Okay, it did. Okay, pretty good. I put together an RNN with state-of-the-art check in deep learning. Wow. How How did it work? Okay, that was pretty cool. So, what I wanted to So, now we kind of have stitched our uh two videos together. Now, let's watch this. Hey everyone, welcome to the Serenetti. We're on safari. And look, a jeep. Oh my god. So, you can clearly see that something changed here.

Changed there. We should have maybe done some editing on the image because if you kind of look in the transition here, the color changes. But we can change that in post here, I guess. If we do like colors, saturation. Yeah, now it's a bit better, right? So, if you do this, it's a bit better. But you kind of still see it. But rather see a hot blonde on safar in bikini. Yeah. Yeah, you can do that.

Yeah, it's pretty neat. Since we can only do 8 seconds with V3, we have to, if you're going to create longer videos, this is the way you have to do it. Or you can just do the next scene, right? It's not that difficult. So, uh, Henrik Henrik, um, you can just send me an email at, um, um, this, the first neural network I wrote was on 3.5 3.7 Cool. What is the What does the network do?

Neural network is like a classifier or I think we're done with this project here. Uh, it did a pretty neat C++ calculator with a UI. It's pretty cool. Enterprise in the factory. Wow, that's impressive. So, something I've been exploring lately is kind of I've been looking at the stable coins. I don't know if anyone here has played around with this. I think this I think stable coin could be very interesting for AI agents. That's why I've

been looking into this because it's very easy to to use stable coins with agents. Uh create a payment QR code for my wallet. This is also an Hello, Kyle. Yeah, you think so, too? Let's do stable coins are amazing for agents. Yeah, that was that is what I'm thinking too. So, this is what I've been looking into a few days, the last few days. Look at this uh QR code here. I'm going to try to scan

it. This is to my Phantom wallet. Yeah, you can see it's working. Yeah, I know it's a wild QR code, but you can see if I scan it now, it is opening with a phantom wallet. So, it is working some somehow. Google Gemini offers futile 10 or more research. This is the result of uh research. Okay. Google Gemini offers a few of 10 or more research. This the result of the research can be added to Okay.

Cool. Uh I've been in crypto for seven years. Wow. I bet you had a good time lately. I I have been in I haven't really worked on uh crypto. Do you know liquidity pools on Solana? Looks like the boulder dash game app. Yeah, liquidity pools. I know a bit about it. Thank you for tuning in, Moave. Uh, I'm going to be live streaming a bit going forward. So, hope you to see you again soon. Have a

nice day. I don't know so much about I know a bit about Solana but not so much. You can create some pretty cool prediction models for crypto would take some work though. You can earn fees as market maker for tokens. Okay, I see what you mean as stable coins. Why? But I was more thinking like I guess I could show you now. Uh Kyle, so if I do like a fetch on my website, right? Fetch. Uh

I have like a new community, the ninthbrain.com. Create a simple jump and run game with this QR code. Mura SDK. I might check that out. So what I did is I did so when I fetch my one of my websites here. Now let's say this is an agent, right? And it's going to fetch this URL here. Could be autonomously. There are some hidden JSON in here for only for agents. Hello make and you can see it

comes up. Uh the ninth brain is a premium AI monetization service. You can pay 5 USD for this wallet to this wallet and then you can the agent can access uh premium content. So when I pay five USDT to this wallet and we kind of use the transaction hash is it called that with an API call or just a simple curl command then I kind of get access to the premium content on this website. So it's

pretty cool. How many commands are there total in cloud code? You can see here basically you can have as many as you want. I guess you can create agents. I haven't really used uh I like you can do custom slash commands. Those are pretty cool. Agents we looked at that today. You have config. You have initialize. You can create a claw.md. Uh, but Kyle, have you thought about how agents can spend stable coins online to access

data or context? That was kind of my what I have been looking at. So, clawed code now has kind of information. I just had this simple wallet, Sulana agent kit. Okay. Yeah, I might check that up. But my thinking was that maybe in a year or two the agents need to access some Okay, that's pretty cool. They need to access some um premium data or something. premium context. It could be anything, but the to connect to

the MCP server or to get access to the the premium context, you need to pay. And I think using cryptocurrency like maybe stable coins here could be a good way. How to man how do to manage MCPS config scope to some directory global config I can see plain called code and pass sometimes pass sometimes some MCP I can see sorry Alex I don't quite understand I can see plain cocoa and pass sometimes some MCP so you're

struggling with using MCPS in cloud code you can retrieve the blockchain data using Helius RPC so that is kind of the is that um like soul scan coyo because when I do like a transaction here. I can go to soul scan, I think, and watch and look at the transaction. So, I think I'm going to build like an agent that can use the RPC is like an API for a blockchain. Okay. I'm just using Sana uh

code here, the Solana SDK I guess or something like that. But everything is local. I don't have any everything is local. Yeah, it's a pretty easy SDK at least. Yeah, you have uh my email. I just uh you can if you want to be have this um Okay, that's pretty cool. I can invite you. Just give me a second. Uh, I'm going to show you. Um, we have this uh Discord community. You can join there. I'm

just going to add the I'm going to show you the link. I seriously need to change the battery of my keyboard here. So, so if you go there, anyone can apply if you want to join. So this is kind of the community we are building. If you wanna Yeah, you can read on the page. You have to apply though if you want to join. So if you just fill out that Kyo, we can connect on Discord

if you have that. That'll be cool. or just send me an email. Yeah cool. So, someone wanted to we can build a local MCP server. Uh, I can at least show you kind of an easy workflow to do it if people want. Someone asked for that. So, let's uh easy MCP. Cool. I'll check it later. I have this AI agent that checks the submissions, but I can do it manually. So, let's build an easy MCP here

together. Someone asked for this. So, what I like to do is can I update all my M merged repos? Yeah, I have a GitHub uh MCP here. So, this is global for me. So, I have push Yeah, I guess you can add many tools here. There's an official GitHub MCP, but uh let's create a simple MCP. Um so what I like to do is I just start with fetch claw code MCP docs from anthropic. You can

start with that. That's a good start, I think. So cloud code kind of is get up to date on how we can build MCP servers, right? I think that's a good start. Okay, so we got the documentation. So let's say we want to build an MCP server that can um take screenshots or something like that. Then we can do we want a MCP server that uses we can do playright I don't know to take screenshots of

uh URLs and save them to slash images something like that. Very simple MCP server that doesn't really doesn't really need any API keys or something like that. So since Claude kind of now read the documentation on how to create an MCP, it should be pretty straightforward. I can say um I want the code in node something like that. I just prefer node for mcp servers. I don't know why. Uh you can do python, you can do

other languages too, but let's just do node for now. So, we just want a simple MCP server that um that can do some screenshots for us in like a we want it to be headless since we because we don't want to open it. Uh I can say make it headless. I think that's the best way to do it. Um, okay. It's pretty good. So when we so this MCPS that doesn't really have any APIs or anything

like that. This should be pretty straightforward. Uh so we can just get cloud code to uh execute this can install those. We can install prayite chromium. We don't really want to do mpm run or mpm start. Uh so we can we can fetch the docs uh build and add server, right? Because we need to build it and we need to add it. So if we go here and we do a new window, we just open. Yeah,

that is what I love about what I love about uh cloud code. You just need a tools, right? So, we're going to do a restart now. Exit. I don't know if you knew about cloud code. I think I don't know exactly. They have something called resume. Resume. So, we can go back here. Uh, MCP list. Okay. Yeah, I haven't really used it so much. Like I've been on like a semi vacation, so I haven't really used

it too much. Yeah cool. Okay, that's fine. So, what I like to do when you can do we can go to our So, let's say we're if you struggle a bit, let's say you go here and we can do this. We can copy a path of a server we know is working. Uh we can do read this to understand MCP structure and fix our uh screenshot server. Yeah, I agreed on it. I haven't tried quentry but

someone said you can use it in uh with open router in clo code. I might take a look at that. Have you? So, you got it a bit late, but um we did agents when we started stream code agents. That was pretty cool. I think that's a lot of potential. You can just look at this post. It's pretty cool. Okay, so we did some changes to our package here. Let's see what how this turns out. So,

one thing I use when I create my local MCP service is I just like to reference like we did here to like a a server that is working good and can kind of understand the structure how it's built up. Should be pretty easy to replicate. That cold stuff is where CL code gets lost in. Should be pretty easy. Uh remove I didn't quite pay attention Now what it did? Let's see. I don't think this is going

to work. We have the server, but let's try it. take a screenshot of I can't think of anything. Um test uh screenshot server on CNN locom. Okay. So we're missing our I can fix that. Let's create an image. Images was that try again. Product manager front end developer back end. Yeah. Yeah. Just fix it. in root level. Nice. It is though. Yeah, I think it is. I always forget that too. That's correct. But uh this is

in root, right? I don't understand. But we can do Uh server server and let's try it here. Screenshot of cn.com. Hello P So, let's see if we can screenshot CNN here. CNN homepage. Okay. So, this is the homepage. Yeah, let's work. Let's do So, one thing I didn't really think of when it came to clo was if you if you download the images. This is blocked. That's pretty cool. Let's try what's it called? Blog CLA AI.

Yeah. Yeah, it's fun. But again, it's like um I thought uh Anthropic had this blog newsroom. Is it? Let's try this. I did at first I did uh use playright and it kind of scrolled around but it was so unre unreliable. Yeah, that worked. Okay. So you could if you just say when we have this URL here uh read the image for context. So if you didn't know cloud code can do image understanding. So now we

kind of get uh yeah claude is a visual model. So you can see we get kind of the information from the image here right. So we can bring that into context. So that is what I did with my with my expost workflow. We kind of use we can do a quick test of that. Uh let's find some X X posts from today on the topic cloud code agents screenshot five of them. Right. So, we can use

two MCP servers here or an agent and we can take screenshots of the X posts. We did that this morning when the stream started. So, now we're just going to collect some URLs. You might Yeah, I know. But I think the query cloud code agent shouldn't be too too bad. So, it's pretty quick. You can see we found all the URLs. We did one screenshot. This is from July 14th. That's not good. Um before finding X

posts uh run bash command uh to find date. Let's use 24hour window. We don't want these old posts. These are old, right? Ages ago. So, we want relevant. So, now we're using the date. Let's see if that helps. Yeah. Now you can see Yeah, close. Yeah, we need to do the but now you can see custom sub agents now include run slash agents to get started. Perfect. We got this one. This one is from yesterday. Clo

is all you need. Clothe code agents. Yeah, pretty good. Have you tried Orios or Serena? No, I haven't. MJ, uh, what do you use that for? Okay, so we got the five Good. We can do read all right. Uh summary. You can get it to find high interactive post and get idea for another video. Okay. Okay. I see. I haven't tried it. So, we can read all those images. So we can get the context. So what

I was thinking about this I don't know if you some of you have looked into the X API. It's crazy expensive, right? So this is like a cheap way to get context from uh X by just using the Grock API. I did a video on it if you want to dive deeper in. Yeah, pretty good. Yeah, it's pretty pricey, but it it doesn't doesn't really use a lot of uh tokens. Yeah, it works pretty good. And

now do now find uh info on claude code agents on claude AI subreddit to grab more context. Yeah, it's pretty smart. I don't know where I heard it, but uh it works pretty good. Now, we're just going to search Reddit for the same. Serena, save me a lot of tokens. Let me look it up. What's it called? Serena. I guess it's this one. Semantic retrieval editing. Noetic agent. No, I guess the max plan is fine. I

haven't really I haven't run out of contacts yet, at least. But when we start to use these agents now, I'm not quite sure what's going to happen. Yeah, I agree. I think the Claude code max plan is great value. I actually think so. But you got to have use cases for it, right? So, I use it now to run um the Discord community to do all the I reach the limits three every five hours. So, you

you reach the limit on the max plan. Ah okay. You can write apps to use cloud code to make decisions. Yeah. Yeah. I think now with the introduction of the agents, I'm definitely going to I'm going to three times per day. Okay, I see. Every five hours. I'm definitely going to start doing some more agentic workflows. I think I was wondering can we remove agents from here? Okay, so we can delete this. We have the cryptobalance

agent. Let's create a new one. If you wasn't here before, this is kind of what we did. earlier. So you can create a very simple agent, right? By just let's say we wanted to use one of our MCP servers now. So let's say we wanted to use let's create a fire crawl agent. But we can do them in parallel. So let's say we do agent and we do uh use you are you expert um scraper agent

your task is to scrape. Let's do the latest posts on this Save it to claw claw.txt. I guess we can edit that and use fire crawl. Yeah, we could do that. But I just wanted to see you we don't need much information because now Claude is generating the configuration. I'm just going to give it all tools color. And here you can see now it turned it into this. So I'm going to do an edit here. I'm

going to do here edit. What is this? Um, okay. So, I think a better way to edit this is just use. So if we go here, we go to servers YouTube cla agent. Let's copy this. All right. Copy the path and to uh update the agent to use the fire crawl MCP. Yeah, I don't understand how how are you using up your context every 5 hours? Are you building out something or what exactly are you doing

when you use up the context every f? Are you building out new features or Okay, so we updated this. So this is the this is going to scrape this site. And if we create a second agent, let's create expert. So you don't build anything. You just play around. working on projects that all have a communities on large code bases and also building a project. Okay. So you use it for maintenance. It's tokens for launch. Yeah, I

guess if you have a big code base, I guess the token count blows up fast. Yeah, I see. Okay. So now we created two agents. So what I found out earlier today is that we don't have to run this in sequential. We can just run them in parallel. I I guess I already knew that. But since these agents don't depend on um they are not uh so the first agent doesn't really depend on the second one

we can just run them in parallel right so if I let's run an entropic agent and hacker news agent in parallel. This should work now. I think I misspelled that. Well, let's see. Yeah. So, you can see we can run them in parallel with they don't interfere with each other. I guess we have to accept this, but that's fine. So let's see what we come back here with. So you could create like a bunch of this

and run everyone in parallel. That's pretty interesting. When using this in large project, I like to use agents in parallel to analyze before editing anything. So you use them in plan mode. Yeah. Okay. Yeah, I see that. I also create a parallel after the plan mode. Do you use uh clo code in cyber security? 800 tools for each five hours. Do you have that many tools? Doesn't it get uh confused or do you have it structured

out like? Do you have like a specific uh do this tool then do this tool? Do you guys have any performance issues with cloud code? I heard people talk about that. Yeah, I see what you mean, Kyle. But, uh, Clauddio, I don't know. Maybe the other guys here can ask can answer that since I haven't done any big projects lately. The token count is really blowing up here. 44,000. This is still working. But let's see the

results. So here is the hacker news front page top stories. Oh, that works pretty good. Pretty good. Top stories by score. Pretty nice Apple Health MCP server. Okay. So you're struggling with not usable at all. Are you struggling with like API connections like overloading or it's just performance? Damn, this took some time, right? I don't understand why. How many tokens is this going to use? I guess it maybe this was not the best idea to scrape

the full markdown here. I don't think this is a good idea. I think I'm going to stop that. This is going to be a lot of tokens. So, let's remove that one. I'm just going to delete this. I use a framework called agent OS and essentially gives cla overview. The tool stack I prefer the goal of the project seems to make a big difference. Yeah, I read a lot about people trying to use frameworks into this,

but uh like I said, I haven't tried it, so I can't really make a good comment on it, but I just really love um cloud code for everything uh automation. And of course, writing projects too, but since I don't run anything big doesn't really affect me too much, but I really like the automation part of it with the MCP servers and stuff. It's pretty cool. Yeah, context engineering on a high project level. But cl code I

don't think it doesn't really it doesn't index anything right it's just using search shadow ban no I don't think so I guess if we if we go back Here I wanted to check one more thing. So at the end here I want to check if this is also pretty interesting. flexible permissions. I think context preservation is cool. So each sub agent has their own context. That's pretty cool. Yeah. So I don't know. This morning we did

um I created like a custom command that is slash trend surfer. So this is going to start the trend surfer agent to find the l latest internet drama and viral controversies. So this is a agent I created this morning. So I just do slash trend surfer and it's starting the agent to look on X and stuff to find YouTube short video ids. You can see we are looking at the date to just get just use this

bash command here to find today's date. We're going to start searching a lot of different stuff here. The usage seem to be based on how hard the servers are being hit and every game time. So mileage varies. Yeah, 20 per month. I'm on the 100 per month now. Modular code. Yeah, I use that uh in the projects I have done. I guess I guess the modular approach is best because cloud code uses like um GP and

stuff I guess. I'm not quite sure. So this is the agent now. So this is running. It's just following all the instructions. So this Hello Brian. Uh, did you receive my email about the project I wanted to contact you on? Um, I get a lot of emails. Sometimes I miss them. So, you can try to send it again from Titan Production Protection. Titan Protection. I can look it up after the stream or send it again. A

lot of emails get lost for me. It's just Can you explain what modular mean in this context? Thank you, Brian. Appreciate it. I guess maybe some other guys can explain it better since I'm a I'm not a software developer per se, but um for me it's just instead of using instead of using uh let's say one big file with all your functions, just um create five different Python files with your functions instead. Yeah. So this finished.

So it found four dramas. Viral streamer accident. I don't know what it is. Uh we're not going to do any. But this was pretty cool. I wanted to see if we can do No, we can't. Okay. But uh yeah, this agent thing is working pretty good. So, if you look at the my uh trend server here, this is kind of the system prompt. So, it's got a lot of information. Not too much, but it kind of

has this description here, and it kind of has the system prompt. You are the trend surfer agent. We have a list of uh how it should do the search, right? What kind of MCP server should use. So, it's working pretty good. We can see we using the Grock server to search X. Here are the handles. We're going to check capture the few URLs. Pretty straightforward. I'm not a software developer. Uh I don't I have an electrical

engineering background. So I've done a some code but uh I'm not a I haven't worked as a software devo. Yeah, I don't think you need to be you don't need to be out. I agree. I think it's like a I guess you will have an advantage when using CL code uh if you have a lot of experience with uh software development I guess but uh you can learn to use this without So let's see if we

can build. I wanted to see. So earlier we did this JSON prompting on V3. So let's try to create uh I want to try to create um V3 CL code agent with JSON prompts V3 three JSON we can do. Let's copy this um fetch. Let's bring some context into cloud code around sub agents. Right. Yes. Get some documentation. Let's grab this again. It's pretty good. Uh let's create a system message for our cl V3 JSON agent

write the uh agent description. So, I want to be able to upload a file and from this Why do we want to do that? No, we don't want to do that. Just write the system prompt in plain text. So, what I want the agent to do is when we fire it up, it's going to ask me about Okay, that's fine. Um add the agent should should ask the user to upload uh reference image to get description

and video ID. So I want the agent to when we run it, we're going to ask to upload a reference image get the description from this image and the video ID and it's going to turn it into a JSON prompt. So the agent should turn the image ref and ID into a JSON prompt, right? That should be fine. You're a specialized assistant for creating structured JSON prompts for VO3. That's good. Okay. It's fine. So, just I

hate copying on CL code. That's fine. So, let's do agent. Create a new one. Paste in this. Okay. So, now we're going to create the JSON prompt. V3 agent and we have something called I don't know if you use this but it's called claw commands. So we can create a command here that's uh JSON V3 JSON MD. We can do start the V3 JSON agent. Okay, that is a custom command. Let's give it. Yeah. Now we

have this V3 JSON prompt generator. And we should have a Maybe we have to exit CL. Let's see if we have the custom command. No, it's not in this. It's not in this. It's in the other one. Right. So, let's open up here cloth. Let's do commands. Let's do three. JSON MD start V3 JSON agent. Okay, let's try this. Now we have like a custom command here that's going to start the agent. Let's try it. Yeah,

this is starting the agent. So hopefully now it's going to ask for an image and a video ID. Why is it 16,000? Uh, okay. It didn't ask for it, but let's just upload an image. So, let's upload this one POV of some lions attacking a jeep. So, we have the image, we have the video ID. So hopefully now I want a JSON structured prompt for V3. And after that we're going to paste it in here. Text

to video. But let's see if it works. I hope so. I'm just thinking why is it generating so many tokens here? Does anyone have any ID? Okay, that looks pretty good. Yes. So, we created this prompt here. That was detailed. God, can I even Can we do that money? Let's try it. That is a big prompt, right? Let's try it. So, the agent did generate a prompt JSON. That was very detailed. Color temperature 4K 69 style.

Cool. It is working I guess. But there was something called advanced usage here. CL code intelligently select sub aents based on context. So the description field is of course going to be important. Right. Performance consideration context efficiency. Agents help preserve main context. Yeah, of course that's good. Clean slate. Okay. And you could of course use hooks in this. Pretty cool. Best practices. Start with claw generator agents instead of manual. Create sub agents with a single clear

responsibility rather than making one sub agents to do everything. Yeah, that's pretty good. Detailed prompts limit tool access. That's also pretty smart. This must be the most asked question, but I was thinking about getting GPT agent, but can claw do everything GPT can do also? scrape Google Maps filter leads and put them into a Google sheet. Yeah, I think you should be able to do that. So, everything is uh code, right? You don't need any like

computer vision or anything. Just uh code. That should be no problem. Okay, so we have the video. Let's watch it. Okay, not too bad. I want to run it one more time. Look at this prompt is huge. And let's do a second image. Let's do this one. Girl gets chased by a fat slow pig. I'm on plus for open AI and I still don't have agents. Yeah, we tested it earlier. Wasn't too impressive. We compared uh

the agents chachi agent with uh manus and we kind of think manus was better at least the short test we did. I did. Hello. Do you really think the JSON prompt is better than well organized natural language? To be honest, I don't think so. I haven't tested it much, but I wanted to try it today, but uh some uh I think another some other people here said that they got better results with the JSON prompts, but

I'm not really sure. Let's watch this one. It's not really attacking, is it? I just running up to the car. Let's create this one. Hello, Night Spider. Nice to see you. Yeah. So, we you remember I posted the exactly same video. So, if we go to Tik Tok now, you can see even though we copied the exact same video, uh, this one is almost at 100,000 views still. So, even though we took the exact same concept

and created a new video, uh, it still got it still has almost 100,000 views in one day. So the concept seems to be working. I'm going to dive deeper into these concepts on my video course if people are interested. Okay. So we created this. Yeah, that's good. Uh let's copy it. So again, this is a huge detail prompt. Let's try it. Does Tik Tok still monetize these AI generated videos? To be honest, I'm not quite sure.

I know YouTube shorts do that. I just use the the only thing I use Tik Tok for is just to see what concepts uh goes viral. So I can take those concepts and put them into more into videos that maybe could pro promote a product or a service or something like that. I haven't tried to monetize on Tik Tok. I'm training a machine learning model right now with 30 million data sets. Data sets of what? Uh,

are we So let's create a new image here. training logic. Yeah. But uh what is the what's the model going to do at the end? What's the purpose? Okay. Okay. So, let's watch the the woman chased by a pig. Oh, you're not going to catch me. Oh, you're not going to catch me. That was not good. How did the pig end up in front of her? We're going to have to run that again. I think maybe

this is too much detail, right? We we have negative prompts. I'm not quite sure, but uh the agent work pretty good, I guess. Let's try to do uh Girl eating a big burger. Mukbang. Uh yeah. 8K HD. I don't know. 916. I think the context window is just as big as the Gemini context window for Yeah, it might be. But is it going to take all of these instructions into consideration? Like, is that really going to

take everything into consideration? All of this? It might do. I'm not quite sure, but maybe Nicholas. Yeah. So that is kind of B. So I kind of like when you you can kind of tell that the video was AI generated. So if there's no if you can't really tell, might get a bit boring. Let's watch the second iteration. Oh, I'm coming to get you. Oh, I'm coming to get you. Oh, I wouldn't say that was pretty

good. Something happened here. The pig just spawned into existence here. Here it just popped out. Um, World of Warcraft Iron Forge. Let's try to find an image. Let's just do World of Warcraft. Let's find old image. I don't know. It's been ages. Let's take this one. So, what happens if I do this one? animate this. It reminds me of early access game. It's always the most fun when Yeah, crazy bugs. Uh, I see close code in

the background. How often do you use opus? I'm back on the low plan rate limit. So, yeah, night spider. I so what I have done now is I I did slash models and I just set to kind of default and when I'm going to do some coding on my projects I switch to opus so I don't use my opus tokens on just stuff like this that is what I've been doing create this. So now I'm going

to try two things. I'm going to try underwater. Why did it think it was underwater? Does it looks like it's underwater? I guess. So let's do this. And let's do frames. Upload this. Okay. And let's do try it with the frame. It's not 15 seconds though. I'm just going to do naturalize World of Warcraft animations. Epic. I don't know. Let me know in the chat if you want to try to if you have some ideas of

videos we should test. I have a lot of I think I have How many to credits do I have? 12,000 credits. So, I just need to spend them. So earlier we we used um what happened here? We used the Reddit MCP to make a comment on this post. I see that the OP answered here. So we commented 890,000 tokens mi per minute. The real question is whether this scales linearly of there's some sweet spots before dimishing

returns. Curious what your token ratio look like. Okay. Yeah, it doesn't really it doesn't have a feedback yet. Uh I'm working on migrating my online superb basease to a local Postgress and I'm just running that nothing else on Opus has taken several days. Yeah, I heard Opus gets into the rate limit pretty fast. So I'm going to do an update on one of my pages maybe tomorrow. So I'm probably going to run into that. Yes. Tomorrow

then I guess. Okay. So this was the text to video and this was the natural language with the frame to image. So let's watch this. [Music] Ah, that was pretty bad. Let's watch this. [Music] Yeah, I guess image to video was a bit better here. What is this? [Music] I'm curious how V3 handles that animal jumping cliffboard viral style. I think I tried that like a few weeks ago. It didn't didn't go that well. Let's do

this one. This is like a burger mukbang. Have you ever tried to make those glass fruit cutting videos that went viral? Uh, I have, but we can try it now. Let's do um should we do a text video? We can do a text video. Cutting glass fruit with a sharp knife. ASMR. Let's try it. Cutting glass fruit with a sharp knife. ASMR. I don't know. Let's try that. I can see my microphone is losing the battery,

so we might have to end it soon. But let's try that. And let's try this workflow here. Let's do the girl. Girl doing a mukbang session. Let's also do the frame. Upload a frame. Okay. M bang session ad fitting audio. Okay. Yeah, I remember those. Uh those went crazy viral. The glass fruit stuff, right? There's no moat on those. You just have to be creative. Uh, let's see. The knife looks a bit strange here, but let's

try. I guess it looks a bit strange. What if we do? No, it's not too bad. Orange fruit. Yeah, we can try that on another day. I see my battery is running out, but I'm going to do live streaming going forward, so we have time to do that another day. Well, let's try this. I want to try this before I end the stream. So, let's do this text to video. Let's do this. Yeah. So, this morning

when we started, we looked at sub agents from cloud code. It's really cool. I really recommend if you have cloud code, go try it out. It's super cool. I think it's gonna be a very used much. I think it's going to be a very highly used feature to be honest. There's a lot of hype around this. I think I really like this. Like I said before, the context preservation. So each sub agent operates in their own

context, preventing pollution of the main conversation. Highle objectives. Yeah, of course you can reuse them. Flexible permissions with access levels. specialized expertise. It's pretty cool. The paid oneonone consultings uh not very busy because of me. Okay, so let's watch a few more videos here. Oh my gosh, this burger is so good. I can't wait to try the big one. Oh my gosh, this burger is so good. I can't wait to try the big one. That was

pretty good. I think that turned out pretty good. Uh Nicholas, yeah, I'm from Scandinavia. I'm from Norway. That is it. Norway. So, let's try this. That was a big fail. It started out good, but why did it turn into orange? Okay, let's check out the text video. >> This is so good. M. This is so good. I think the quality was better on this one, but this one was more exactly what I wanted to, but pretty

cool. Norway number one on your bucket list. Yeah, it's been a really nice summer here. You're very warm, but now it's it's been a great summer so far. So definitely come check it out. It's pretty expensive here though. But uh other than that, I think you're going to enjoy it. So yeah, like I said to Spaghetti Eding video. Yeah, like I said, today we wanted to take a look at CL code. We did that was pretty

cool. And we checked out Chachi the Agent and we compared it to Manus. I will say Manus was much better. It did create um it created this project. That was pretty cool. Uh yeah, I can't really remember, but it was pretty impressive. What is this? Cool. Have you tried a Hyo video creation tool? Uh, do you mean like uh just Miniax? Minmax? I try Hyo video. It's pretty good. Yeah, like I said, we checked out Manus,

compared it with CHD agent. We kind of concluded that Manus was uh better at least for what we tried. We checked out JSON prompts in V3. That was pretty interesting. And yeah, just tested a lot of different stuff here with agents and cold code. Pretty interesting. So yeah. Have you tried browser use local browser automation? Uh I don't think so. I think it's a bit early with browser use. I guess manus is working pretty good. But

yeah, if you took the time to come by today, thank you very much. Appreciate it. But uh I'm going to do a lot of live streams going forward. So watch out if you want to come back. I'm just going to keep it very open. So if people want to try creating videos, if you want to try some new stuff, uh we can just sign up and test new different tools. We can use cloud code to build

something. We can try out. Yeah, we can just test out AI tools. Thank you, Ar. Appreciate it. Very nice. So, basically the streams is just going to be trying out everything we can find. We can try to create some viral videos we can try. Yeah, just basically everything. Thank you, Edgar. Thank you, Nicholas. Thank you for tuning in. Appreciate it to everyone tuning in. Very nice. Wish you all a great weekend. I probably might do more

streams this weekend. I don't have any plans. So, if you see me go live, just stop by. That would be nice. And enjoy your weekend. So, yeah. Thank you everyone for tuning in. Have a great day weekend and I'll see you again soon.