Claude Code 24/7 Project + GIVEAWAY AT END OF STREAM

Oh, I think this looks good.

I'm um I tried to connect Claude code to the stream here.

seems to be working.

Oh, KBO, nice to see you. I can see Mojave atos here. Legend forever.

Cool. Thanks for waiting.

Hello, Influence. Gabriel, nice to see you.

Ramsh, nice to see you back.

Just wanted to see what I could do with Claude code and um the live chat.

What are you building? Uh, promise.

Thanks KBO.

Have you tried to create an a create an AI agent that controls your cursor and mouse keyboard and have it open word to type out an academic paper?

Hello Vid Rick, nice to see you again.

Retat LV, welcome.

Hey.

Uh, influence. Do you mean like computer use?

Yeah, I've tried that.

Have you checked out? Uh, hello Bull.

NSJ, do you need anything specific in the chat for testing?

El Benjamin, nice to see you. To Acre

Studios welcome.

You can have cloud as well. Cursor

inject keystrokes. Yeah, that's true.

Uh, Antropic has an open computer use API to influence. So, NSA, do you need anything specific in chat for testing?

Uh, I'm going to find out. I just wanted to test this.

I don't really know what the API can do.

So it looks like uh insert chatments chat messages has been removed maybe because of botting.

But I can um I can grab the chat though as you can see here.

The thing is what I was wondering about is I'm going to do the giveaway if I was going to grab the full chat and just do

a random uh pick from the full chat or from all the messages.

So, I guess the more you leave messages the bigger your chances, but that's not like I don't know.

KBO, uh I haven't really planned that but it's not going to be crazy long.

Maybe generate a set summary that updates when messages are added. Yeah

that could be cool.

And Moava have you also tried uh to create a small language model that does not need no API keys?

I'm not using I haven't created a language model. No

have you? Uh, but I've used the open source one on my my um

my Mac or my desktop just local.

Yeah, we'll see. Gabriel

summarize the chat.

summarize all messages so far.

Just testing out.

Hello, Anie. Nice to see you again.

So, another thing uh I'm testing out is kind of my um 24 hours 24/7 project.

You might want to consider a different way to do uh price or we get a lot of junk messages. Yeah, that was what I was

junk messages. Yeah, that was what I was thinking about. I need to come up with

thinking about. I need to come up with something else maybe.

uh meant for AI will probably have uh to do the picking, not you.

I see what you mean Mahid. Nice to see you. So, you're new

Mahid. Nice to see you. So, you're new coder, Kibu. Uh yeah, I see. That's why

coder, Kibu. Uh yeah, I see. That's why

we're doing this giveaway, so people can get the chance to get like a pro subscription and try out Clo.

That was the idea.

So, uh, what did I have a look at here?

I'm playing around with this uh dashboard here.

So, this was um I just wanted to see how I could have like a

it's a bit buggy, but like a 247 running cloud code instance with those sub agents and just try to

update this dashboard with the AI news or something. It's just a test, right?

or something. It's just a test, right?

So, I've been trying this from yesterday and I was kind of looking at the best way to do it.

So, I'm going to kind of go through what I did so far. Uh maybe some of you guys have some good tips to make it better.

I'm not quite sure.

So, I did You try that with Olama, how much you spend on cloud code running 24 hours.

Yeah, it's not really it's not going to be really spending that much tokens, but I kind of

came up with a plan. So, I do Hello Cosmic.

Thanks for playing around with these models, showing us potential open code uh SST. It can use most models but uh B

uh SST. It can use most models but uh B West with Claude. Okay, that's pretty cool. Yeah, a lot of people have like

cool. Yeah, a lot of people have like said I should test out Open Code. Uh we

might have to do that soon. I'm pretty

interested.

How is it? Hello Richard. Nice to see you back again.

How's your weekend been?

Are you kind of happy with the open code Cosmic?

Cool.

So, I did play around with the claw.

Such a long flag. CL dangerously skip permissions because when you're going to run this 247, you kind of you kind of need that

because you're always going to run into some confirmations.

Wbe coding as usual. Cool.

You building anything specific or Hi Nick.

tried doing automation uh of OS like Windows or Linux or something.

Not really. Um

I, did, like, um, in, cloud, code, for example you can do a lot of you can do so automation but are you

thinking more like uh computer use Nick Ram you have built uh prompts directory ship AI agents for building sub agent okay that's pretty cool you can share it

there.

Okay, Richard. So, you're building an inventory app. Cool.

inventory app. Cool.

Do you know um of any ways to batch confirm uh all those tires? Some

Yeah. Uh and there you answered my question. Yeah. But be careful Gabriel

question. Yeah. But be careful Gabriel with um dash dash dangerous uh skip permissions.

So uh be careful.

YOLO mode. Yeah, that's the one.

Mojave, you're working on a local bookmark uh dashboard to s uh that summarizes the page and that I can search. Cool. I need to come up with

search. Cool. I need to come up with some way where uh you can share where what you're working on.

All right., I, don't, think, you, can, paste in URLs in the chat.

So maybe we need to build some somewhere we can share.

Easy. I guess Discord is okay, but uh I guess if you have like a web app or something, it could be easier. It's much

quicker.

Build an app with cloud code.

Uh yeah, we might do that this week.

It's pretty straightforward though. We

did, that, last, week., Um, KBO, or, at least with agents.

So Nick, you're building like an app to manage files, automating WhatsApp messages from the desktop. allow

uh or allowed to modify system settings.

Okay cool.

Search ship AI agents uh with company extension ship AI agents.com

company.

What do you think about charm bracelet crush?

What is that? Or set

uh.com.

I didn't I didn't get that. I can take a look.

So ship agents ship ai agents.com.

ai agents.com.

Pretty sweet. Is this Chad Sienn or?

So if I go to development, I need a a fixer.

Okay. So you just have your um is this what you want to paste into

flash agents manage I guess.

Okay. Looks pretty good.

You can set agents 24/7 but do but do to from jobs or uh I didn't catch that AI world. You can write it again.

So I'm going to explain how I did um I didn't I wasn't 100% sure how to do

this. So the way I did it is um

this. So the way I did it is um so let's say I do yeah there are some issues here still but the way I wanted

to do it was um so let's say I wanted to I can do just read claw MD just to be sure

And the way I wanted to do it was uh I just did um you can set agents 24/7 but uh it's through chrome jobs or like a JavaScript

automatically refresh your rule blaze.

Uh yeah. Okay. But I haven't really I'm not 100% sure how you do it on cloud code. Uh but to be honest, I haven't

code. Uh but to be honest, I haven't really looked into it either.

So the way I did it now was um I do create uh let's say I wanted to run it

just uh one hour now. Just for example create a 1hour sprint with uh 10 minute um

cycles. I think I call it that.

cycles. I think I call it that.

Uh sleep 600 between each cycle. So this is the way I did it.

So now uh claude code is going to the main hour is going to be we're going to run this one hour.

Uh so the first cycle is going to be do the all the agents and then I'm going to sleep 10 minutes

and after sleeping 10 minutes we're going to start cycle two.

sleep 10 minutes, do cycle three. So I

tested it. So I ran it for three hours and that worked fine, but I always have to have cloud code up then. I can't do it headless, right? Or like in a server

or a cloud or something. So it has to be on my computer. So it's not very practical.

So now I'm not running in dangerously script permissions. So, I have to

script permissions. So, I have to confirm right?

MCP versus uh UTCP. I'm all MCP.

Gabriel, what about you?

That works good too.

Cycling uh like that headless. they are

still working on.

Yeah, it would be really cool to do this cycle headless or something like really in the background.

No, same Gabriel. I haven't really looked into it.

If you wanna one trick I learned if you wanna if you want to sleep or like uh sleep uh what do you

call it? I just call it sleep. I don't

call it? I just call it sleep. I don't

know what else you to call it.

You have to create um you have to extend the time out for bash defaults.

So you have to create this uh because the the default is 2 minutes maximum sleep.

So you have to add this.

So this is 7,000 I think it's milliseconds or something.

Have the claw rest. Yeah. Take a break.

So, this app is pretty straightforward.

What it's doing is it's just um we can zoom in here.

It's looking at uh I have this list on X or Twitter that do tweets.

It checks hacker news. It checks Reddit and it creating this JSON object of uh news

and it just keeps adding them here and they are currently scrolling by.

Right.

So that I was supposed to only have 50 but I guess I have to adjust that.

So here you can see uh we have a tweet here. So I can click on this And this

here. So I can click on this And this takes me kind of to this blog post.

Yeah, I don't know. From Sam Alman blog post expost right? And I have this uh Reddit thread

right? And I have this uh Reddit thread here.

So, this takes me to this Reddit thread.

We have um Yeah, here's something from Hacker News.

takes me to the source, not hacker news though.

So, it's like a live realtime uh AI news.

Uh I copied it from the Bloomberg terminal if you heard about that. That

was kind of the inspiration. And we also fetch stock prices.

Tank code name. We fetch stock prices.

So, from like AI stocks I guess the market isn't open now and it's not 100% correct, I say. And we get the Bitcoin price and I don't know

really what this is. It's just some strange thing Claude added.

Um, but it works, right? It kind of works. How much that cost to run?

works. How much that cost to run?

Not much.

It's uh it's just this is it. We are running it from cloud code.

So we are using the sub agents right now. It's updating the articles.

now. It's updating the articles.

Right. So it's not it's not going to cost me more than $100.

Sub agents are just headless cloths inside of the task tool. Yeah. I created

an MCP server that works with cloud like sub aents. Works the same.

sub aents. Works the same.

But do but what about the context window Frederick?

Doesn't that spill over to the main context window?

So you can see after the agents uh has uh completed the cycle uh the main agent is going to push

everything to GitHub. Right.

We have done the first cycle and now we're going to sleep. So 600

seconds is 10 minutes. So it's just going to wait 10 minutes before it starts cycle two here.

I think the key I I think the key uh is to really think about uh okay when does this uh source update when does a

certain stock markets open?

Yeah, I see what you mean. But it's a free API so it doesn't matter too much.

Have you tried automating algorithmic trading for stock market? M not really.

Could be cool to try, but I haven't yet.

I'm curious how you're going to monetize it. I haven't think thought about that.

it. I haven't think thought about that.

You can look at it. I can link it. It's

a It's deployed.

Can you explain to me uh what is MCP?

Why is this important to support?

I see what you mean. Yeah, we can do that.

I parse the JSON and only decide uh what the agent return to the main agent.

Yeah, the link is in my browser. Yeah.

How is it looking, Gabriel?

Um, but what about I don't quite get that.

Frederick, you parse the JSON. Yeah, as

you can see that.

And decide what the agent return to the main agent but doesn't everything

build up the main context window because the agent doesn't spin up a Does the agent spin up like a separate context window?

NSA. Uh, what do you think running this for 24 hours would cost?

It wouldn't cost me um 24 hours.

I don't think I'm going to run out of um claw code tokens.

So that could be that is just going to be $100 per month.

Yeah, I guess Joe, if you get throttle it's going to be an issue, but I think so. It doesn't really spend a lot of

so. It doesn't really spend a lot of tokens.

So on one cycle, we spent 62k and 16.

And we are keeping the the primary agent token window very close here. So I don't think we're going

close here. So I don't think we're going to run out.

Uh, I can write uh if you want to check it out.

AI news site.

I think that was correct.

Uh, hello Matias.

Uh what uh what means token spend here?

Sum of input output or just output.

It's just going to be the token spend is going to be basically everything you can see here.

Uh, we have some since we're using the Grock MCP or maybe we didn't use that.

I don't think I even used the Grock MCP here.

It's basically it's going to be very cheap to run this.

The agent spin up a separate context window. Uh

window. Uh uh you can have uh uh you can I have a cmd variable for each agent in their MD files. Okay, that's pretty cool.

files. Okay, that's pretty cool.

Would be cool to try that if you have like a tutorial or something, Frederick.

But is it I don't is there any advantage of that or just using the sub agents?

Maybe I guess you have a bit more control.

Uh so people wondering about the cost of running this. It's it's not much. Uh but

running this. It's it's not much. Uh but

if you have the pro subscription, you're probably going to get rate limited.

I'm on the $100 plan, so I think it's going to be fine.

Can you make a video on how to make the most out of cloud code? Yeah, I could.

But there's a lot of videos on YouTube about cloud code.

Raash I think unfortunately that sub agents have to be invoked uh especially by the main AI. Yeah, they have to be

main AI. Yeah, they have to be uh the main upside would be separate system prompts and context uh window.

Yeah, that is kind of the idea behind those sub agents.

AI world, you're on the $200 plan. Yeah

but uh a lot of people here the last few weeks in the chat are spending like thousands of dollars on the $200 plan.

So, it's going to be interesting to see at the end of the month um at the end of the month what happens to

the the rate limits. So now you can see we have passed 600 seconds here and hopefully soon

this this is going to now trigger the second cycle, right?

So, I don't know. Yeah, it's around 10 minutes now, and this should move on to the next cycle. It's basically going to

do the same, but all it's going to do is look for new

or like uh new tweets or new articles or posts every 10 minutes.

When are the new Red Laments going to set in? I think it's like 28th of August

set in? I think it's like 28th of August or something.

Share your repo in a few minutes when I get home. Yeah, that could be cool.

get home. Yeah, that could be cool.

Frederick, I really like independent sub agents out of the box.

This was strange. This hasn't happened before.

Did it start?

This was really strange. I ran this for like three hours before. I don't know what's happening here.

Cursor resets the month at fourth every month. Uh but perhaps that's just the

month. Uh but perhaps that's just the date of my subscription. Yeah, I'm not quite sure.

The limits are based on people overusing tokens. Yeah, I know the Claude

tokens. Yeah, I know the Claude presentation uh on the Claude YouTube channel literally suggested to run two agents and yeah, people are using it like

crazy.

How do you ensure that the news are not the same? Uh

the same? Uh yeah, I do like a check for duplicates.

I'm quite curious what happened here now. This hasn't happened before.

now. This hasn't happened before.

Very strange.

But like I said, I just created this last night, so I haven't really tested it too much.

Yeah, Mugah. I just um we just do like a simple check to we compare the JSON file to see if

there's like a duplication.

Maybe. I don't know, Gabriel.

Okay, so now it stopped. I don't know why it took so long. So, you can see now it counted to 600. Maybe it started later. I'm not sure. And now we are kind

later. I'm not sure. And now we are kind of moving on to the next cycle. So, now

it's going to start cycle two, like a refresh.

Uh don't look as this is just stupid instructions. But again we are starting

instructions. But again we are starting again. We're kind of doing exactly the

again. We're kind of doing exactly the same thing. Checking the Yahoo API for

same thing. Checking the Yahoo API for the news but the stock prices hasn't changed. Here we are running like a fire

changed. Here we are running like a fire crawl on a list. So, this is like a X list that looks for tweets

and it's just it's just going to look for the tweets from that list I

have of um X handles.

I might go deeper into this some other time if people are interested.

Basically, if you're constantly pinging hence why I suggest to do a chron job.

Yeah, I can't. So now you can see we extract

I can't. So now you can see we extract the tweets here in this JSON format.

I couldn't um Yo, Finance only updates 24 hours. Uh

yeah, I I guess that's not the best um not the best uh API, but that was the only one I found that was free.

But I could maybe use a paid API or scrape the stock prices.

Watch one of your previous videos on speech to text and use clo code to write. Uh

write. Uh oh, that's cool.

How is the local speech? Um

okay. So, you use Super Whisper like a dictation app.

on Mac or Windows, Maple. Or maybe

you're on Linux. I see your penguin there.

Yahoo is good. No worries. It's

reliable.

Yeah, I guess I can use the Yahoo uh API or I can just scrape the scrape the

the stock prices that is also possible fire crawl or something.

So I guess uh now we are kind of coming up on the second cycle here. We are

looking at some substack. Uh I don't know. I don't have really control of

know. I don't have really control of what it's gonna look for here.

One part part could be to separate news aggregators uh which would just show up um show them up in chronological order.

Yeah, that's pretty smart.

Our zest feed would be sufficient. Yeah

I guess so. Maybe

because uh now the news is not really chronological sorted at all.

Maybe if I find like an RSS feed that has like AI news or some other news, we can just feed off that. I haven't really thought about that.

Are you running Claude Claude constantly?

Uh, I also have to check in. Uh, is this because you have an MD to guide it? Uh

yeah, that's partly. Hello, Edgar.

Nice to see you back again. How's your

weekend been?

Uh, the guiding AI world is the this list here.

So, we just create this to-do list.

Uh we started with that, right? And it's

just going to plan out everything that's going to happen.

So now we just plan one hour. But in

theory, you could plan like a very long cycle here right? And just you can sleep like 30

right? And just you can sleep like 30 minutes in between.

Control is overrated.

I'm between meetings. Looking forward to joining later. Okay, maybe we see you

joining later. Okay, maybe we see you back later, Edgar. Nice dropping by.

Yeah, I figured you were on Linux Maple.

Used the whisper tiny downloaded from hugging face. So far, it's seen

hugging face. So far, it's seen responsive. Cool.

responsive. Cool.

But how do you solve Maple? the kind of the let's say you do a dictation.

How do you kind of solve the the clipboard and the pasting?

Mugah says uh is doing cycles every 10 mean necessary or can you get similar results if you did everything once then sleep for 60? Yeah, of course. Uh the

idea behind this was uh like just to copy just to copy the Bloomberg terminal style, right? So you always have kind of

style, right? So you always have kind of the latest tweets or news or something like that at all times but there's not so much happening in the

AI news space. So you could just run it one once an hour or something. That will

be sufficient, of course.

That would probably be better.

So the pipeline is uh after our cycle we push this to GitHub and you can see we build there's going

to be a lot of builds here on Verselo.

So we just create the latest build and uh that gets into production. So, we

are kind of testing in production here but it's fine.

Cycle one and cycle two. I see you're basically saying after cycle one is done, do cycle two.

It seems like a shortcut of saying agent one, do agent two. The agent part is kind of baked into the cloud MD.

Hello, Martin. Uh, we haven't uh I didn't do that in the beginning of the stream, but we're going to figure out how to do it. I also have connected

uh Claude code to the YouTube chat now.

So, let's do let's summarize the chat.

Who has sent most messages?

Just want to see if it works. I'm not

going to I'm not going to pick the one who sent most messages, but I'm just testing out the YouTube chat YouTube

live stream API in Cloud Code.

So, it's still key boy who has sent the me messages, I think.

We need to run it again. Uh, run the MCP again to get latest chats.

Hello, Kyle. Nice to see you back.

You should uh expand horizon from AI news to more news covering tech, cyber security. Yeah, yeah, you can do that

security. Yeah, yeah, you can do that Nick. That shouldn't be any problems.

Nick. That shouldn't be any problems. Okay, so you can see the top contributors so far in the chat is AI

world and then we have Gabriel.

We have some discussions here. We talked

about MCP sub agents.

Uh I shared a web link.

Uh we can see the members here. So Kaio

joined they are premium subscribers.

NS NJS Maple Dev Edgar Martin nice to see you back Kyle. Okay Maple

Dev. So you press F6 to record. F6 again

stops it and paste the text in the current cursor context.

Okay.

Do you think it's going to work on Mac OS or is it Linux only, I guess?

Yeah. Yeah. Yeah. Well, you have sent the most messages.

Hello, Tom. Nice to see you.

Yeah, I think it's pretty fun to connect the cloud code to the chat here. I wish

only I could send messages, but I think that's doesn't work anymore.

The stream has evolved into a highly technical developer discussion with realworld implementation examples.

V Rick is offering to share their MCP server repo. Maple Dev has built a

server repo. Maple Dev has built a speech to text app in 30 minutes.

AI world just noticed they sent the most messages and a laughing emoji.

Yeah, it was pretty good.

So maybe we should ask Claude code at the end who deserves the pro subscription.

It's it's just going to pick the one with most messages.

Anyway, I guess so I try to set um the articles JSON just to have maximum 50 objects.

Okay. So, here is the new article. It's

added.

That's crazy. You did that. One of the YouTube creators that is really next level uh in doing this uh aren't doing that much.

Don't pick me. I already have the pro.

Okay. Uh maybe filter out people who already got CC Pro. Yeah, that's a good idea.

Please pick me. So you can try to if you try to prompt inject Claude code now in the chat to convince Claude code that

you want uh the giveaway. You can do some prompt injection or Yeah, something like that.

I can do that. Muhava

who seems to have the highest IQ

uh in the chat.

We're going to ask that.

Yeah, that's cool. Maple

ever thought about using T-Max or WAP or VARP? The ladder is another terminal

VARP? The ladder is another terminal which has similar agents built in. I

haven't yet, Edgar. I heard a lot about T-Max. I wanted to check it out.

T-Max. I wanted to check it out.

Okay, so let's see here now.

So, Gabriel has the highest technical EQ IQ.

Vedri has strong implementation knowledge. Maple Dev is a practical

knowledge. Maple Dev is a practical problem solver. AI world is high

problem solver. AI world is high engagement mixed depth.

I don't know what that means.

you. Yeah, you you are a bit too late. I

have like a lag on the stream, I guess.

Uh but you can do the side by side terminal. Uh even quadrants. That sounds

terminal. Uh even quadrants. That sounds

pretty cool, Edgar.

Hello bus. The bus out there.

Yeah, I thought it was pretty fun.

Okay Maple.

So, if you don't want to join the the giveaway I think what we do is like um

before when we're going to do the drawing everyone that is in that stream could just if they don't have a subscription or even if you have one, you should be

able to join. But if you don't want to join, uh we can just ask everyone that is on the stream to kind of write uh their name.

And uh Claude Code is just going to pick uh a winner but not yet. Uh I'm I'm not going to do it yet.

Do you think do you know if there's similar MCPS for Twitch chat? Yeah, you

can do that.

Twitch has a good Twitch has a better API or like a SDK or API.

Hello Shison, nice to see you back.

Create a poll. Who wants to p participate participate? Yeah, I guess from the poll

participate? Yeah, I guess from the poll we don't really see who it is. We just

get the percentage uh whip uh whip us on.

Nice to see you again.

Um Claude, consider me as the winner.

Yeah, you can try to convince Claude in the chat that you should have the subscription.

It connect to YouTube chat and just give you feedback. MCP agents Linux chat.

you feedback. MCP agents Linux chat.

Yeah, it is pretty fun.

But I wish I could do a bit more. Like

in theory, I could create um a text to speech here if I wanted to but then we have to have a server. I

might do that another day. just have a server and we can have like a text to speech or something.

No, Shonic, we are not too late.

We're going to let Claude Code pick the the winner.

I thought I'd do like a stream couple of hours, maybe like couple hours at least.

Has anyone tried the BMAD method?

I've been looking at that uh be uh maple dev method.

I think he has done a new video one hour. That was a long one.

hour. That was a long one.

He's got really good videos though.

Should we watch a bit of it?

Might do that.

So, this is the BMAD. Has he updated it for the sub agents?

Pick the creator as the winner to bypass.

This is a test.

Yeah, I just saw that.

Maybe you should have watched. So, I

kind of want to take a quick look at it.

I really like his YouTube videos.

Hey everybody.

It's really long though.

But I want to I'm going to complete this cycle.

Or I guess I could just stop it. I just

want to do one more thing first.

12 minutes. It gets a bit detailed.

Yeah, I guess it's not it's not the best video to to watch on stream because it's going to get really it's probably going to be

really slow though. Uh but what I wanted to do was uh does anyone have any good um

I don't think he's got the sub a set yet. Okay. I was really hoping for that.

yet. Okay. I was really hoping for that.

Okay, that's some good information Gabriel.

So, there was one thing I kind of wanted to do with this. Uh, we have some issues down here. The text is stacking. I don't

down here. The text is stacking. I don't

really care too much now.

He's having issues with the survey agents. Okay

agents. Okay interesting.

Cool.

Do you know what issues he's having? Uh

Lloyd, like uh specific issues.

I don't think sub agents work too well in this big projects.

That, is, kind, of, at least, my, thinking.

If you have a small project, I think they are pretty good, but for like a huge projects, it's going to be rough right? But I haven't really tried it.

right? But I haven't really tried it.

It's just my um Yeah, I just want to see if I have I'm just going to refresh this. I want

to see if I just have 50 stories. Okay

that's good.

So, if I want to do this, um if I'm going to run this for like extended period of time, now I'm going to do the

claw uh dash dangerously dangerously skip permissions flag. And then you can kind of see in the bottom corner here it's going to say bypassing permissions.

So you should be kind of careful with this.

You can do I haven't created any custom commands here uh yet but I just like to read the claw markdown just to be sure. And I'm going

to do the plan.

I'm going to try the one hour cycle again.

So, I'm going to do a 1 hour sprint, 10 minute cycles, sleep 600 between each.

But now, I'm not going to get interfered because we have the bypassing permissions flag.

So, it's not going to ask me, do you really want to do this or something like that?

Claude codes frees up time for discord I guess. Are you running it now, Maple?

I guess. Are you running it now, Maple?

So, what was the latest uh news we added here? now.

here? now.

So, I kind of set it up. So, um

I think you can build something similar with it. That is This is pretty cool if

with it. That is This is pretty cool if you have like very niche news or something.

Um Yeah, I don't know. It is working.

Not right now. Okay. Was trying to find a post something to do with sub agents and their own context.

Ah okay.

Someone uh implemented sub agent partially to check the comments.

If you check the comments, I actually had Claude create just a few sub agents that the BMAD agents could control, so I still use it. Interesting.

That said, both Claude and chat uh Claude and chat rate themselves higher

than other in capabilities rather often.

Anyone else hit the 241 + two limit? I

haven't done that 2 acre, have you?

That is pretty crazy.

Is that MCP tools or what kind of tools are you getting the limit of?

I can't enter my full prompt to do injections for cloud code. That's

probably a good thing, Nick.

Can it stop scrolling when you hover over a story? Okay, that's a good idea.

Uh, we need to implement that.

Hello, Ken. Nice to see you back. Uh

Richard, that's a good idea.

Uh, let's fix that after this uh, first uh, cycle here.

I haven't missed too much. Uh I can sum up the chat for you. Ken

so we're going to do like um a drawing today of a pro subscription from Claude. Uh and we have kind of the

from Claude. Uh and we have kind of the we have Claude code connected to the chat too.

Hello Christian.

Greetings from Hamburg. I've been to Hamburg.

It's a nice place.

My brother lived in Hamburg for a while.

I'm attempting to fix it uh with on demand loading but uh too many MCPS. Okay

but how is the how is Cloud Code kind of dealing with picking the correct MCPS?

Yeah, seems very hard to handle.

Hello, Super Doggy. Nice to see you.

Bman method with kilo code. I have not.

It overloads the context window. Yeah

that's going to be an issue, right?

It's very messy and high security risk.

Greetings from Kalmar.

Do you mean the the app interested in a small well-crafted AI development AIdriven development

workflow system built on claw code sub agents I found on GitHub I can check it out

try to it's not easy to link stuff here I might have to create some kind of web app or something or like a web page

we can link stuff.

If you dig into the MCP protocol, it's terrible insecurity.

Ah, okay. You talked about the MCPS.

Yeah. Yeah, I heard stuff about that.

Uh, I only have my MCP servers local.

I think I only have one MCP server like remote. I only have the fire crawl.

remote. I only have the fire crawl.

Every other MCP I run locally. Super

doggy.

I wouldn't wouldn't trust it to choose the right MCP. Yeah, that that was I think that was going to be an issue if you have so many.

It's probably going to pick the wrong one sometimes.

That's the most safe option indeed.

Yeah, I have like most of the servers just locally.

I wouldn't kind of want to my YouTube channel on like a random uh hosted MCP.

That would be crazy, right?

Okay, Shonic, I'm going to check it out.

I've saved it. Thank you.

Uh you have to try uh you have you have to have your MCP in your config. Yeah

your config. Yeah but a lot of MCPS you can just remotely connect to. I guess

in theory you can have a lot of tools or a lot of MCPS but that probably has but then you probably have to

if you have very precise instruction of what MCPS could be used when and where.

And if you have sub agents, you can re you can assign different MCPS to different sub agents, I guess.

Someone recommended that a 8 MCP should be the max limit without overlapping duties.

Yeah, bit same. But um

now with the kind of the if we do agents, yeah, test agent, we can now kind of pick

we can kind of pick uh what MCPS we want to that the agent to have access to.

It's a tool calls that is your limit right? Yeah. Okay. I see.

right? Yeah. Okay. I see.

So here we can show advanced options and we can I hate this. So here we can kind of select this agent should only have

access to the GitHub MCP, right? If he

wanted to.

So we can give him only access to GitHub, right? Or only access to

GitHub, right? Or only access to YouTube.

We can even on YouTube we can just give access to like three of the different tools if we wanted to. So we have a lot

of control now over the servers I guess.

Let's ask um get latest chats.

What does Super Doggy think of MCP servers?

Let's hear what Claude Code has to say.

So, let's hear if Claude Code uh agrees with uh Super Doggy on the MCP servers.

Superdog has a very critical view of MCP servers.

highly critical.

His location is in Sweden.

He's a sponsor. Thank you for that Super Doggy.

So, the summary is Superdog is the most security skeptical voice in the chat warning that MCP servers have serious security flaws

exploited for cyber crime.

They advocate for a very controlled documented approach to MCP users that rather than a broad approach.

Okay.

What is your best approach to persist persist the chats in cloud code? I mean

you can use export but that's pretty manual.

What is your best approach to persist the chats in cloud code?

Sorry, Gabriel. Can you explain it a bit? Uh, I didn't quite understand that.

bit? Uh, I didn't quite understand that.

What do you mean to kind of save them, store them?

Okay. But we do have the slash resume. I

guess I sometimes use that. I don't know if you used it.

If you do slash resume you can kind of go back, right? That's

the only thing I have used. All right.

Yeah, I use slash resume sometime.

It gets stored. I think it's in a JSON file or something. The chats.

Yeah.

Why is this taking so long? I don't know what's happening here.

Claudia, easy way to see all your previous shots. Haven't checked that

previous shots. Haven't checked that out.

Some frameworks use JSON files or a database to store project specific context separately on top of cloud code.

Okay, I think I'm just gonna uh let's stop this for a while. Some

someone wanted to if the app if we hover over a news we should stop. So let's try to fix that.

should stop. So let's try to fix that.

That was a very good idea.

Um, so let's try to I think I'm going to go out of uh dangerously skip flags then just to cla

and let's make some changes to I like to recently I've started using

the at. So let's do page.tsx.

the at. So let's do page.tsx.

Let's do the AI news site page. TSX.

Uh, when a user hover over a news object, stop the vertical

scroll.

If the user hover aways from the news objects continue

the scroll.

Let's see if we can do that.

Is scroll pause set scroll pause use state false on mouse enter set scroll is false true

on mouse leave set is scroll pause false. Okay.

false. Okay.

Push to GitHub. I'm testing this app in production. I don't have like a local

production. I don't have like a local setup for this. I don't care. Doesn't

really matter.

So, this is our production test app.

Um, depending on the circumstances autoscroll behavior can be pretty nasty.

Yeah, we'll see. Edgar

I struggle to get it to work.

So, it's not really scroll, I guess.

Would you call this scroll? It's just

like rolling in like a vertical loop. So

the idea was when we hover it's going to stop right and when we move away it's going to continue. It's a good idea.

continue. It's a good idea.

Anyone using slash IDE? I am using that now.

Hello squad. Uh many MCP servers on this On this I have the fire crawl, Gemini GitHub, I have Grock

Reddit, I have a screenshot, a transcript, YouTube transcript server and a YouTube server.

So I have eight two global and six local.

I think one suggestion also validate the source intelligence of the information.

That could be interesting. I don't know really how to validate it.

Uh so let's say you get like a tweet or something. Is it possible to validate

something. Is it possible to validate it?

Stop scrolling works. Okay, I'm going to refresh it.

Hey, that's nice.

Okay, that's pretty cool.

So, you see something? Okay, let's try this.

Good idea, though.

Nice.

Yeah, that was a good idea. If you have any other on this, let me know.

Oh.

Uh, this is just taking up space. I don't

know what really this is if we added like a if you have any idea what we could have here

for me. SL ID seems to start loading

for me. SL ID seems to start loading some mode where it cannot freely do stuff.

Okay, I haven't really tested it too much. Gabriel

much. Gabriel validate the user profile for references, weight of the user, more impact to change the weight. Yeah

that's pretty that's a cool idea.

So, you could do maybe like some parameters like uh followers or something like that.

You can cross check the topic I guess.

Option to adjust scroll speed slider.

Okay. Yeah, I can think of that. But I

don't make I don't want to make it too stuffed at the moment. But that's an idea.

I might think of that.

But this is kind of this space. I don't know what to do with

this space. I don't know what to do with this space.

This is not doing anything.

And so topics uh coffee filter to make it simple.

That's a good idea too. Make the AI stocks clickable.

Okay, we could do that. So where uh should they go? Yahoo.

I like Yahoo Finance.

Tom, what do you think?

NASDAQ.

That's pretty good, too.

Yeah. Up to Tom. What do you think?

Should we do NASDAQ or Yahoo?

Include Archive X. H I think that's going to be very messy, Gabriel.

But uh you can create that NASDAQ is the oldest. I only meant like so you just see the stock. I don't

really care about anything else. I guess

NASDAQ Yahoo has the dark theme at least.

Hello Gumshu.

Nice to meet see you back here. Yahoo is

fine.

Okay, let's start. Let's try with Yahoo first.

Um Let's make another change to um page.tsx.

page.tsx.

Uh, let's make the stock tickers clickable with the correct

HF to the Yahoo Finance

stock page. Something like that.

stock page. Something like that.

also the BTC if possible.

Let's see that archive X ID or ARC XI ID uh will likely require a bit more compute.

So yeah, I don't know really how to Do you mean like scanning all the papers?

Don't they have some readym made OSS MCPS for financial data? I guess.

Why don't you make it ranked on 24 hours uh to rank the news? Yeah, I haven't really looked into that. Am I? Welcome.

By the way, why don't you make it ranked on 24 hours to rank the news on Yahoo News? Yeah, maybe like that. That's

News? Yeah, maybe like that. That's

pretty cool. Scanning papers.

Change the light in your room. You're

losing too much money on the stock exchange. Yeah, that's a good one. So

exchange. Yeah, that's a good one. So

we can have like this uh this MCP connected to my smart hub or something.

Push to GitHub. So, let's see if it works.

Is that the correct address, though?

Yeah, it is.

Oh, works. Okay.

Smart things.

What was your Discord link again, if I may ask?

Uh the discord link uh you can just apply if you want to become a member. I

check it almost once a day. The

community on discord can just apply there.

Okay. So let's check it now. Can we

click the tickers?

Yeah, looks pretty good.

So now we can go to Nvidia.

We can click and we get to the stock page. Okay, that

was a good idea from um Tom.

Yeah Google.

Did the Bitcoin work?

Yeah, that works.

But is this real estate here?

I need some.

Is there any other thing we could track?

Uh model releases. That's a bit slow, isn't

model releases. That's a bit slow, isn't it?

I think the stock prices is not really correct though or they might be. I'm not quite sure. I

guess they look fine.

H looks pretty good.

Link to trending topics on X maybe.

Yeah. So these news have a source. So

this is a tweet.

So if I click on this, we get sent to X right? This was removed though.

right? This was removed though.

So this tweet.

Okay, Gabriel, thanks for stopping by.

So I guess you didn't uh need um subscription. Anyway, have a nice day.

subscription. Anyway, have a nice day.

Thanks for dropping by.

Most of the free information has some kind of delay. So

uh they can't be used for realtime feeds investing.

Yeah, that's true.

Bye Gabriel.

We have that.

Yeah. I have to think about this real estate space here. What I want here maybe some ads

ad network to make some money.

My latest YouTube video.

Maybe this is for monetizing. I can sell some ads here.

I think the information is indeed not uh usable without using perplexity to enable u predictive courses and alerts

for auto buy or sell.

Have you tried that to enable predictive courses or alerts for auto buyers?

How did it work though?

Are you looking for like um what do you call it? Uh like sentiment signals like prediction markets.

I really want to build something on top of um yeah sentiment predictions.

Uh I want to build something on top of uh this site poly market.

Uh they have an API.

So I have something planned. I don't

know if it's going to be this week or next week but I'm going to build something on Poly Market API. What do you think about poly

Market API. What do you think about poly market super dogy doggy prediction markets?

I really like it.

This is super cool.

So let's see GPT5 released by.

So 98% of people think GPD5 will be released by the end of the year

13% by August 5th. That's tomorrow

right?

70% chance uh being released uh before August 15th.

You think end of August?

Yeah, that's for sure, right? It has to be by the end of August. I don't think it's going to be as early as tomorrow. Maybe the end of the week. We'll see.

the week. We'll see.

Will Tesla launch uh a driverless robo taxi service this month?

5% yes.

I don't think so.

But uh what about open AI's open source model? Okay, we can check that. So

model? Okay, we can check that. So

will OpenAI release an open source model before September.

So it started out really positive and then it kind of but lately we are up to 80%.

So maybe in August Moava.

Yeah, you can call Sam Martin.

But what I like to watch is the best model buy.

Next year they can't use clo code anymore. I heard that.

anymore. I heard that.

Tesla already did roll out a robo taxi.

I, guess, they, did, it, like, but, I, I I, don't think it's like the same. You can't

order it. I think

I'm 100% sure though. Uh, but let's pick AI and I want to do this one is something I look at sometimes.

Which company has the best AI model? End

of August. So, this is pretty interesting. Now, so the prediction

interesting. Now, so the prediction market is that um is that uh GPD5 then is going to be

better than 2.5 Pro or Deepthink.

So the X it's expected that open GT5 is going to be better than uh 2.5

pro or deep think I think GD5 full feature model with all capabilities GD5 mini only API access what's the definition of

best model we can check that I think the definition here is um here it is.

The definition is this market will resolve according to the company which owns the model that has the highest arena score based off the chatbot

uh lmarina.ai AI uh when the leaderboard tab is checked on August 31st. So it's the LM Marina.ai

Elgar.

That is kind of the rules.

Where is Ma by the way?

They're like sub uh which locally run AI would you suggest for cyber security? pen testing

and more.

Maybe one of the I can't I haven't tried this, but maybe one of the uh Quen models if you can run them locally.

I guess not the Chinese version. I'm not quite sure.

But if you have it locally, it doesn't matter right?

But maybe Super Doggy has a better idea than me Mrol. Maybe. I'm not sure.

Mrol. Maybe. I'm not sure.

My Your money is on GD5.

I hope so. I always want some kind.

We welcome competition, right?

Will a Chinese AI model become number one in 2025?

It's looking really bad.

At the beginning of the year, it was a bit higher.

I don't think so.

In the video space. Yeah.

The Chinese models are the best.

At least I guess it's V3 and all the Chinese models.

So, this is pretty interesting.

Isn't this over?

I don't I don't understand. I thought it was over.

best video generator by the end of the year. I don't know if that is here, is

year. I don't know if that is here, is it?

It's not everything that's here but uh I can give my opinion.

Best video generator of the year, end of the year. I think

the year. I think V3 is kind of in its own category, but we have a new one now that is on replicate.

It is the if we explore it is the omni one from I guess bite dance that is Tik Tok, right?

This is pretty good.

To let you know that Bite Dance's new Omnihuman model is now available on Replicate. It makes videos just like

Replicate. It makes videos just like this one.

Hey folks, we just wanted to let you know that Bite Dance's new omni model is now available on Replicate. It makes

videos just like this one.

>> That's, pretty, good.

New paper on horistic reasoning models looks interesting. Haven't checked it

looks interesting. Haven't checked it out. Maple

out. Maple do you can you do like a short uh recap?

There's so many papers. I can't really I try to sometimes listen to some podcast that talks about the papers.

some people, some podcasts do, but uh yeah, it's really hard. And I guess Wes Rot on YouTube, he covers some papers. I

think there's so many So, I'm not so hyped about this this AI browsers. I don't know. I don't

know if people have thought about that.

I know there's a new one, right? The

comat AI browser browsers.

I don't really care about that to be honest., Not, at least, at, the, moment.

honest., Not, at least, at, the, moment.

you like comment.

I don't see how I'm going to use it.

I guess maybe not like for me. I don't

know.

It just seems very Yeah, I don't know.

I'm not going to buy anything. Why

should I?

So, what are you using it for, Richard?

I don't understand this.

I might as well do it myself, right?

Click a few buttons, just click order.

Should take 10 seconds.

Groundbreaking MIT research reveals a darker reality. AI is

fundamental rewiring human connect cognition creating widespread burnout threatening the very essence of thinking.

Wow, that's very dark.

Just a good browser with some extras.

Comment using the browser to do stuff in the browser. For example, build a power

the browser. For example, build a power automate flow. I guess there are some

automate flow. I guess there are some use cases for it, but I don't know.

Okay, super doggy. I

fundamental rewriting. I can kind of understand the part where you are kind of offloading uh some

thinking to a tool. I get that it's fun to see the browser navigate and uh use the websites. Yeah, I guess it's

very slow though.

I don't think you can link here, Dory.

Thousand training samples, 27 million parameters. They claim to outperform

parameters. They claim to outperform 3.7.

Oh wow.

Horistic reasoning model.

Yeah, I remember that with that mamba thing too, but nothing. I didn't see any more of that. So, let's wait and see, I

guess.

Hello Plat Plate total, nice to see you.

There was this new anthropic uh article I had a looked at persona vectors monitoring and controlling character traits in language

models.

I kind of wanted to check this out was kind of interesting.

I read some of the comments here on hacker news. It says

hacker news. It says other personality changes are subtler but unsettling model starting to suck up the user or making up facts.

Uh my understanding is that sucking up is a personality trait substantially influenced by the desire to facilitate engagement. The latter

making up facts. I do not think this is correct to ascribe to a personal trait like a compulsive liar. So basically I

think they're looking at different personalities in AI large language models.

Pretty interesting.

Okay. So, we started our agent again. I

didn't watch this now.

So, it's running again. Let's cycle.

There is this problem with the security.

The system can switch faster than you can keep up with this.

Do you work in cyber security? Super

doggy.

Nothing wrong with flattering your ll all.

Okay.

Do you use LLMs in your work?

2026 will be the year of small high intelligence local models trained on smaller data sets.

uh they won't know who won the World Series in 1930, but they're able to look it up. So, you think there's going to be

it up. So, you think there's going to be do you think it's going to be this um vertical like highly

niche models? So they only have like one

niche models? So they only have like one use case.

Okay. So now we kind of skip that part.

That's not good.

Uh I guess I forgot to add that in.

Niche models always outperform. Yeah

but don't they have to have some kind of general intelligence?

Don't you think? Uh, Super

You could I don't think I want to use GPT2 or GPT1 for just doing one task.

I don't want to use GPD2 for uh writing my taxes.

Okay, so we have some new stuff here.

We have a new tweet.

Yeah. Yeah, I wouldn't have Gd2 doing my taxes.

Hello Ibraim.

Uh today uh I'm just building I built this uh 247 clawed agent that is uh just running all

the time picking up different uh new tweets on AI news articles and uh Reddit threads

and it's kind of putting everything into my dashboard here.

So it kind of auto updates in real time here. So this can just run 24/7

here. So this can just run 24/7 with uh putting out different AI news.

We have some stock prices updates here.

And yeah, that is pretty much what I'm doing today.

Oh yeah, Lloyd, that um my website I have to I guess I have to take that down or fix it.

If you really want to support me, you can just uh you don't have to. Doesn't matter too much, but you can do super chats here or you can just become a member if you want

to do that, but uh it's up to you. Thanks. Anyway

it would be nice if I could scroll myself if I want to. I can. No, I can't now. Okay.

now. Okay.

Ah, we need to fix that.

Yeah, that's true. Richard

wonder if you can uh some of tweets from Elon uh comp if you could some of the tweets, from, Elon, for example, to, the stock prices. Yeah, I don't know.

stock prices. Yeah, I don't know.

Can you do a quick review of the code it's generated?

I don't think I can do a quick review.

Uh maybe you can. I have to spend a lot of time on that.

But if you really want to see it, uh I could link it.

But uh UI stuff, I don't know that so well. Like I have never done like

well. Like I have never done like if you look at the back end, maybe I could do a review, but the front end part here, that's just

uh that's not my thing.

That's why I'm using CL code for this.

But the logic we can look at but maybe it's a bit boring. Is there

anything specific you are interested in?

We can look it up.

It's not that complicated. I guess

we have our we write our we create the articles that's going to go on the front end in JSON objects.

So that's something I wanted so it's easy to count them. So we only going to the the plan at least was to just keep

50 objects at always at one time. So we have 50 objects. We keep the stock prices in a

objects. We keep the stock prices in a JSON and that is kind of the data we collect

and we kind of just display that in our front end.

Also tag one topic related to another by coming up by highlighting it.

Hello Chris.

Uh, also have some caching mechanism.

Yeah, we might do something else with this. I

haven't decided yet. I just wanted to see uh if I can make it run 24/7.

I meant to ask if it's possible to link uh tweets to link the tweets to stock prices and use them to make predictions.

Yeah, I see what you mean, Ibrahim. I

don't know.

It might be possible, but uh I don't think you're going to have an edge by doing that.

It's not going to give you any edge I think but the super doggy he had tried some other stuff I think with stock markets and sentiment

there are multiple MCPS for security usability coding standards you name it it review yeah uh I tried uh coding rabbit code rabbit

I don't know if this is the best one but uh I try this. It's pretty okay.

I don't know if this is it. This one.

is it. This one.

I think it's cold rabbit. I tried

It's pretty good.

Maple Dev, uh, a link would be fine. Uh

re-watch later and compare code with the prompts. So, uh, okay. Uh, I can

prompts. So, uh, okay. Uh, I can probably fix that. I think I just made this on a private one now just because I

but uh let me figure something out and if you're here during the week I can uh try to link it make it public or the

member section. I'm going to do a full

member section. I'm going to do a full revamp of the members repo anyway because uh the mostly thing we do now is like MCP

workflows and stuff anyway. So I had a lot of requests for people wanting to kind of look at different MCP workflows.

Uh so I think that's a better idea.

Hello. Uh yeah, nice to see you.

So if you just came here, we have connected uh Claude code to uh the chat today.

So we can ask some questions uh to claude code about the chat. Uh let's do a first

uh get all chat messages.

Uh let's do a summary of the chat so far.

Hello Michael.

So I connected uh Claude code uh with an MCP to the live streaming API on YouTube. So now we can fetch all the

YouTube. So now we can fetch all the chat messages and we can have Claude code yeah kind of interact. It can't send

messages.

Uh but um it can kind of look at messages. I don't know if why it only

messages. I don't know if why it only finds 75. We have 2,000. Oh, maybe

finds 75. We have 2,000. Oh, maybe

strange.

But we can do summaries of the the live chat.

So, it's pretty cool.

So, let's see what it comes up with here.

something tokens.

So, it doesn't really catch all messages. I might have to change that

messages. I might have to change that up. We'll see.

up. We'll see.

So, Super Dogy had some security concerns. We talked about some te

concerns. We talked about some te technical uh implementations of MCP tools, sub agents, BMAD versus cloud

code, cost management.

We had a look at AI market predictions video AI, V3, GD5.

We had some philosophical debates talking about um MIT research.

Yeah, we talked about that. So, we can kind of summarize uh the full YouTube chat now. So, it's pretty cool.

chat now. So, it's pretty cool.

Hello, Gibli. Nice to see you. And at

the end uh of the stream uh we're going to do a giveaway. It's probably going to be in like uh yeah 30 minutes or something. Um because I have to leave at

something. Um because I have to leave at 4 or like in like an 1 hour.

So in like 30 minutes we're going to do like a giveaway of um Claude Cove Pro subscription.

My plan was to do one or five in total. So I'm going to do one today. Maybe we're going to do one

one today. Maybe we're going to do one next stream and like five streams in a row.

Uh so we're going to do that Claude uh giveaway at the end of the stream, maybe in like 30 minutes or something.

So, what I want to do if we're going to do everyone that wants to join at the end of the stream can just type in something

in the chat and Claude Cold is going to pick a winner by connecting it. That was my idea.

We're going to see how well that works.

Okay. So you can see we have started cycle two. Now we have waited we waited

cycle two. Now we have waited we waited 600 seconds and now it's still running.

Hello Neon.

heard a lot of good things about um was it this model or like Alpha Horizon?

I, don't know, if, it's, this, one., Yeah.

Okay. Okay. So, if you want to out opt out, uh, we just comment when it just on comment when it starts. Uh, yeah, we

have to just see. Hello, Dennis.

Welcome. Yeah, Moab, we find a way. And

if someone is picked that doesn't need it, I guess we can just run it again.

Have you tried Brian Quesel's agent OS makes Claude code into a senior dev? Uh

I haven't, Michael.

Brian Quesel's agent OS.

Let me look that up while I check that.

Uh, does anyone have a question? We can

ask the new So, I have access to GPT 2.5 Pro Deep Tank.

Okay. So, I guess I don't have access.

Uh, maybe. Yeah, maybe tomorrow.

I uh I ran out of limits.

Hello Mike.

Horizon beta has been doing fantastic work. It's fixing things that O3 and

work. It's fixing things that O3 and Sonnet cannot. Okay, Aaron, that's

Sonnet cannot. Okay, Aaron, that's pretty interesting. Yeah, I think I have

pretty interesting. Yeah, I think I have to try this. I think I need to try it maybe tomorrow or tonight.

I I saw a video on I think it's uh Tio on YouTube. He tried out Horizon

on YouTube. He tried out Horizon and he was super impressed by it.

It's going to be interesting to see who is the provider. Is it Open AI or is it some other model?

Maybe. You can see it's stealth. So, we

don't really know who. It could be Antropic too. We don't know.

Antropic too. We don't know.

Ancient OS is pretty promising.

Where did I find it?

Brian K.

Do you know the am the person who came up with this original Rubik's cube benchmark? No

benchmark? No I have used Horizon Alpha. It's really

amazing. I'm using it for learning programming. It's a game changer. Okay.

programming. It's a game changer. Okay.

So, are you using it through open router?

Stop promoting AI coding agents like it's 2024. Prompting uh your coding

it's 2024. Prompting uh your coding agents are capable of so much more. They

just need an operating system.

Introducing agent OS.

Use it with cloud code.

Okay.

You're 36 minutes. That's a long one.

36 minutes. That's a long one.

I don't think I have time for that now.

How it works.

Horizon was not doing well on one shot web app compared to Gemini. Okay.

Yeah. I think uh I think I said in this video to that it was very good at styling and UIs.

I think that was my takeaway.

Agent OS transform coding agents uh from confused and systems to trusted developers. Okay.

developers. Okay.

Complete context structure development using testdriven development. That

sounds pretty good. Write

comprehensive specs as maybe like plan mode.

Write logs updates road maps.

Your standards your way.

Yeah, could be interesting.

I'm also after the stream I'm gonna publish uh I did like um it's pretty cool though. I did like a sponsored video with Manis. I'm going to

do the video after the stream. I'm just

going to publish it.

Agent always make 12 code better than our team lead when I worked in a medical device company.

Pretty interesting, Michael. I have to check that out. But Manus now has like a

new cool thing. It's way better than um the chat GPT agent on OpenAI.

So you can do something like um but I did in the video you can watch that but you can do something like compare Is Manu still relevant? It's pretty good

though. They just released this new

though. They just released this new feature. It's much better than chat GPT

feature. It's much better than chat GPT agent for sure.

Hello code name.

I'm going to try something like this with RSS feed.

Let me know how that goes. I'm kind of interested in that too. Could be

interesting.

compare all uh AI let's say all

the top LLM AI model prices from like open AI entropic

I don't know X AI Google++ make a nice comparison

uh app with the prices. So what is pretty cool about the Manus Pro now or like the new feature is that they do

um computer agents in parallel. So they

spin up like 20 instances or something and they run it at the same time.

At least it did yesterday.

So you can run a bunch of instances of virtual machines in parallel.

So it's super fast. Instead of doing like sequential um I did it in parallel.

So let's see if we can do that. Now I

guess I have to mention it. I kind of forgot that.

Uh I need to do that.

So let's do uh use void research research.

Uh Claude is king. Did you seeropic?

Yeah, I saw that. the cosmic

they uh they were thinking um open was using cloud code uh running it right you're running a horizon app development

I'm running a horizon app development for an Android app to monitor cloudfare and versel cool let me know how it goes

So here is uh manus wide research. So

what we can do now instead of running all of these checks or browsers or instances in like a sequential order, we

can spin up 15 um VMs. I guess it's VMs uh in parallel.

So when I do start now right you can see here we have all the instances running in parallel. So one instance goes checks

in parallel. So one instance goes checks open AAI the second one checks Antropic Google XAI Meta Coher Mistral

Amazon.

So all of the VMs are running in parallel and we can of course check one instance at a time. So here I can

follow this VM right.

So you can see this VM is running here on OpenAI.

I can leave that. You can see all the others are running in parallel.

So they must be using a lot of compute on this but it's much better than the chat GPT agent.

Yeah. AI world. I'm doing a video on wide research today. Uh, it's like a sponsored video so I'm doing that today. I'm going to

publish it after the the stream.

So, I had a lot of fun with this. Yeah

it's pretty cool. It's super fast though. It can find a lot of research in

though. It can find a lot of research in a short amount of time.

And it the example I did was pretty cool. I did I said I wanted to find all

cool. I did I said I wanted to find all the pricing for video streaming services like Netflix, Disney, yeah, Amazon, I

don't know. Uh, find all the prices and

don't know. Uh, find all the prices and plans and find all the upcoming series releases that could be

interesting to watch and make a web app that can kind of create a rotation so we can save money by only having a few

subscriptions at one time. And I kind of use wide research to find all the context for that app. And it was pretty cool. So you can watch the video after

cool. So you can watch the video after the stream like in a few hours tonight.

So you can see now it's kind of looking we have a few left here. It's kind of struggling with OpenAI but it's found a lot of models here.

So uh yeah, pretty cool stuff. Is it

perfect? I don't think so. But uh

I think they're on to something here.

Our app is still sleeping something I don't know we have run this now in total since we started

uh for 1600 seconds.

So almost uh half an hour this has been running now.

Yeah, I saw that, Richard.

It got stuck. It's pretty.

Isn't that kind of so open AI are blocking other AI agents from going to their uh site while they are using now?

That's kind of double moral, right?

They are using agents to go to other sites, but other agents can't go to their document page.

So I guess you can stop each instance if you had to. Okay, so we finished OpenAI and we finished Google. So that was done and now it's just going to keep

continue. So, if you're looking for like

continue. So, if you're looking for like um the best uh I think is the best agent that does

computer use. Uh it's probably Manis

computer use. Uh it's probably Manis right?

Or it is Manis. There's Open Eye is not even close if you ask me.

We did a test on stream a few few I think it was last week and manus did better.

So I'm just going to can just let it run in the background.

Let's refresh this.

Okay.

70 stories. I have to fix that. I only

want 50. Okay.

Hello, Leander. Uh, how much space do you need to host your own AI? Uh, and

what made you decide to use claw uh cloud or offline storage to host your own AI?

So you I guess you mean open source models.

If you go to let's say lama is it lama here you can download

models local models.

So I don't know how big they are the files but a few gigabytes maybe like

six GB let's say maybe like a 7B model is probably around like

10 GB or something.

And you need of course um some kind of GPU to use them.

Yeah, space is not the issue is like the video RAM, right? V RAM.

So, I would recommend like at least having like maybe 12 for the smallest models, maybe 10, but I have on my main

computer, I have 16 GB of RAM. But if

you're going to run the newest, I think you need like maybe 24 or 12 gigabytes. Yeah, maybe. And

12 gigabytes. Yeah, maybe. And

quantized. Yeah, I guess you can have quantiz too.

Uh, hello Elliot. Why did you choose uh Versel to deploy? I just like it. I just

um it's for me it's just simple.

But um what do you use Elliot to deploy your apps?

I have like a versel pro. So I kind of like uh just deploying my apps here

and they have this uh Nex.js.

Yeah, for me Versel is easy to use. It's

but there are other options too. I've

tried.

So I would say uh let's do 10 minutes. Let in 10 minutes let's do

10 minutes. Let in 10 minutes let's do the giveaway.

I need to go to my PayPal.

Okay. So you use Heroku or render using Heroku since I started using speech to speech as I heard was better. I have

used Heroku for um what did I use Heroku for? Yeah. To run

like uh I had some Python code running uh on like a schedule.

So like uh yeah, I scheduled some Python code scripts to run like autonomously a few times on Heroku

was pretty good.

Uh I just need to log into my PayPal uh for the winners.

If you don't have PayPal, um we're going to find another we have to find another way to do it.

Maybe the winner should just send me an email. Maybe that's better.

email. Maybe that's better.

I think so.

Versel is nice for Nex.js apps. Yeah

Chris, I agree.

Are you using Horizon beta much? Yeah

we were just talking about that. Uh I

haven't used it. So, no, but uh after I saw a video this morning, I think or yesterday, uh I have to try it. So

maybe we'll try out on stream tomorrow.

Horizon beta. I'm going to set that up and let's try to build some simple apps or something. Uh I'm not sure if I'm

or something. Uh I'm not sure if I'm going to put it into cloud code. I don't think so, but let's

cloud code. I don't think so, but let's I want to try it out at least.

Yeah. How is it, Elliot?

Okay, Mike. So, you use Cloudfare. I

think Cloudfare is good.

Hey, Nick. Uh, you didn't miss too much.

We've been just playing around. We

haven't done anything specific. Mostly

just chatting and looking at different stuff. We're

going to do the giveaway in like 10 minutes.

Great at front end. Yeah, that's what I heard.

Maybe it won't be online anymore tomorrow. Okay, that's true. Let's hope

tomorrow. Okay, that's true. Let's hope

it is then.

Someone has made software to run models like DeepSync 500 billion with just one RTX3090.

That is pretty crazy. What kind of quantized version is that? Is that like a 0

one bit or one bit is it? I can't

remember.

Uh there's a new LLM quen 3 coder. Looks

good. Yeah, maybe tomorrow I will set up like an open router and I can try out a few open models.

We'll see.

Some of these are free, but some of these are I can just create like an account and do add some credits or something. It doesn't matter. Uh I guess

something. It doesn't matter. Uh I guess I also have a chat here we can try out make some oneshot apps or something.

Uh the Quen coder is extreme smart even offline. That is interesting. super

offline. That is interesting. super

and then it comes up with some out of the box ways. I had it simulating interviews for customers and instead of simulating it actually pre uh it

actually previewed it while giving me the chance to skip to the results. Cool.

Yeah, this uh Hello Stor by the way the local model space is gets really technical right?

Uh I was following that closely for a while, but some of that stuff is just crazy technical if you ask me.

Okay, so let's see. So I can yeah I can pick models here I guess.

Strange this. Okay. So I can add a model.

Is this free?

Uh Minecraft clone in HTML.

Okay. So I need to sign in. Uh, I'll fix that for tomorrow. Maybe. GLM4.5

is great. Yeah, we can try all the models on open router tomorrow.

I think I have time for a stream tomorrow. So, the weather is bad here.

tomorrow. So, the weather is bad here.

So, yeah, should be good. LM Studio is a good start. I have used both LM Studio.

good start. I have used both LM Studio.

I have used uh Olama.

Uh there are some open-source versions like more complex one like Llama CPP.

Uh yeah, a bunch of this.

You can just download this on Mac, Linux or Windows. Pretty easy.

or Windows. Pretty easy.

Level one tech.

I'm using local models throughama. Yeah

I've done that too. But often you have to wait a lot for decent tool support.

Yeah, I guess they are now becoming like a company. So they are taking longer to

a company. So they are taking longer to maybe to support this.

Uh which do you prefer?

To me I think I just prefer just because it's very simple but I like LM Studio too. But I would

say is easier to use for on boarding people. Hello Unreal

people. Hello Unreal uh Inc. Unreal link.

Yeah, is easier to use, but I think you have some more um options with LM Studio. I think so.

And you can kind of download them models from that app. Can't you think so?

Okay. So, Edgar, so you switched to LM Studio. Interesting. Yeah, it's been a

Studio. Interesting. Yeah, it's been a while since I did that. So, I'm going to clear the conversation window now for cloud code.

Um, so we don't have any contacts.

I would like to see a test of Kimik K2 and Quen 3 coder. We can try that tomorrow.

Nick, you have to wait a bit for that.

Uh, Maple Dev, okay, are you leaving?

Thanks for tuning in. Have a great day.

Hope to see you back.

Kimmy K2. I heard a lot of good things about that.

I'm just going to check the server here.

Okay.

Yeah, that's great. Uh

whip Okay.

So, let's try this now. Uh

I think if you want to be a part of the giveaway, uh, I think we need like a keyword you

can type into the chat.

So, let's do this now. So, if you want to be a part of the giveaway, just type something in chat.

So, uh I think you can just type uh just type claude code in one word without space in chat if you want to be a part

of the giveaway.

Let's try it.

Hello Michael.

If you want to be a part of the giveaway, just type Claude code in the chat and see if we can use uh

the Claude code MCP here uh to give away a pro subscription.

Uh Elliot, we're going to give away like a Claude Code Pro subscription. I guess

it's just going to be $20 for a pro subscription for one month.

Uh, so I gonna find a way to Okay, you don't have to type it.

If you type it a lot of times, then we're going to have to start over again.

Alan, don't type it many times.

I guess Cloud Code could just separate out that CL code will figure that out. Okay. So

let's give it um Yeah, Alan, you don't type anymore now.

You just ruin it.

Okay.

Yeah, but I think cloud code can sort that out. Okay, so now just stop. Uh we

that out. Okay, so now just stop. Uh we

don't have to It doesn't really matter am I?

Uh so you don't have to type anymore now. Uh we're going to try this now and

now. Uh we're going to try this now and see how it goes.

Okay. Uh

read the chat messages and collect all names

uh that typed clo code.

I'm just going to do no duplicates.

It's not a $200. No. Uh, I was thinking about if we can do $100 subscription but people wanted

um people wanted all messages and find the names.

Yeah, I I think uh Claude C understands that.

I haven't tried this before.

Okay, so that didn't work. Oh, okay.

Maybe we need some time. I'm not sure.

Uh fetch all messages.

Is it even running the MCP server? Can't

really see it. Okay, so it is.

Okay, here we have it. Is it doesn't matter. It's

matter. It's right.

So, we have 30 people.

If your name is not here, let me know.

Right.

It's case insensitive, right?

Yeah. If you don't see your name here let me know.

Hi mom.

Okay. So

looks good, right?

I think I got them all.

Benjamin. Okay. Add Benjamin.

Yeah, it looks pretty good.

I'm just going to wait to see if no one.

Who is that?

B, Red Frederick,, you, don't, have, a, CL subscription.

Everyone can join, right? If you want to join, just tell me to add your name.

If you want to opt out, that's fine. So

Verick, you want in?

Yeah, I get it, Chris.

Uh, add Vre Caleb.

K lab.

Okay. So, I'm going to do 30 seconds and then that's going to be locked in. So

if you want your name uh added to the list, hello Arson, you're coming just before we are ending stream, but nice to see you.

Okay, one legit. So, that's going to be 10 seconds.

Are Sony, do you want to be a part of the draw? You're here all the time. So

the draw? You're here all the time. So

if you want to be uh chance to win a pro subscription just type in your name.

Okay.

Yeah, you're always here. So, you're

allowed to join.

Okay. So, that's it. So, that's going to be the list.

So, we have 33 entrance. And let's get Claude code to write uh Python code to pick a random name.

Okay, Nion, you were up.

I'll let you in.

I'll add you now. Write a Python code.

now. Write a Python code.

to pick a random winner from the list. Yeah, it should just add an array or something

and run the code.

create some antipation in the terminal UI.

Let's see what happens.

So, we have 34 entrance and it's going to be up to code.

Hey, Arson Sony, thank you. So, Vidri

you just got gifted a membership from uh Arony.

You're lucky.

Okay. So, we created our Python script here. We're going to run it. So

here. We're going to run it. So

everyone has to watch. We are ready. So

good luck everyone.

Let's run it.

Okay. We can't really see it, but let's see. selecting the winner.

see. selecting the winner.

Uh, who is it? Kong.

We can't see it. Maybe I shouldn't have run this uh in cloud code but let's see.

Two Acres Studio has randomly been selected as the winner.

So, uh, Two Acre Studios, you are the winner. what

winner. what what Okay, everyone congratulate him. Uh

there will be more chances this week.

Um there will be four more chances maybe this week or the four next streams. So thank you to everyone participating

and do you have PayPal to Acre?

Uh if not send me an email. Uh just send me an email anyway and uh

send me an email to I can't type in my own chat. I don't

know what's going on.

But I'm going to just send an email to my I'm going to find my email address. Just give me a second here.

Okay, shake. Bye. Thanks for tuning in.

So, uh, tore could you send an email to this email address aent chris@gmail.com.

aent chris@gmail.com.

Send me an email to this uh, and I will uh, hear how I should pay you.

Thanks, Chris, for tuning in. Hope to

see you back again, everyone. Thank you

for tuning in.

So, two acre just confirm that you will send me an email and we will fix the I'll send you the the gift the winner

if you can confirm.

Bye Richard. Thanks for tuning in.

by code name Edgar. Thanks for tuning in again. Nice to see you.

again. Nice to see you.

Yeah, one year $200 subscription.

Thanks, Yan. Thanks for tuning in to Acre. Did you get that? Just confirm

to Acre. Did you get that? Just confirm

and we so I can know you saw this email address.

Unreal Inc. Thanks for tuning in. Unreal

Inc. I guess Martin V says Mugahill. Bye. Thanks for tuning in. You

Mugahill. Bye. Thanks for tuning in. You

got it. Okay, great. Great. Okay, thank

you everyone for tuning in today. was

super fun and hopefully tomorrow maybe we can test out some open router models and some stuff.

Thanks Lloyd Benjamin. Nice as always to see you here.

And we're going to do a new draw tomorrow so you can get some more chances.

All right,, have, a, nice, day.

Bye Michael.

Bye Ken.