NextJS & Mastra, Security Corner, AI-generated stories, AI in spreadsheets, tweet helper agent

hello everybody i'm Shane i'm from Mastra i'm one of the co-founders and welcome to AI Agents Hour today we're going to be talking about a bunch of different things we're going to be talking how to build AI agents with Maestra and Nex.js we're not going to dive into the code so much today as just talk about overall

architecture and how you might structure a project that uses MRA agents with Nex.js and it is going to be a little bit universal you don't have to be using Nex.js i think a lot of the concepts are going to apply and honestly you don't even have to be using MRA to probably think about the architecture of this but I think it's useful to pick two pieces of technology that you know we at least

I've talked to a lot of people that are using and use that as kind of a a baseline for how you might architect a project we are going to have a special guest come in and do an AI security corner so we're going to have Ally join us we're going to be talking to a few hackathon guests and we're also going to be talking to someone from the MRA team

talking a little bit about how we do and how we think about content and our content strategy at Mastra so with that if you are not already you can follow me on X also if you are not already signed up you're running out of time but there's still a little bit of time we do have a master.build hackathon going on right

now submissions are due tomorrow I believe at 8:00 a.m pacific time so there's still time if you want to get a a small project in tonight if you haven't started if you are already working on it you still have time to improve it so get those submissions in if you are not signed up go to master.build you can learn all about it get signed up uh but let's before we jump in I see

we're having some people start to kind of funnel into the live stream this is live so if you are on YouTube if you're on X if you're on LinkedIn you can drop a comment and I'll see most of them and respond to most of them if I can get to them all so you know just like this hey thanks for sending a comment on

X and also this one so please uh yeah let let me know along if you have along the way if you have questions and I will do my best to answer them if you are in the hackathon I would be curious for those actually watching please leave a comment are you how far along are you you feeling good are you feeling like you're going to get the submission in in time or do you think uh

do you think you need some more time because I unfortunately you can have more time but I'm just curious how everyone's feeling i know I've seen some different states of projects we had someone come on uh on Tuesday and show a really great project so we have a few people coming on today i'm interested to see kind of what they're doing and how

you know how things are going for them and then you know we will likely have some people come on tomorrow as well uh and show off some of the things we will be live streaming the awards ceremony so if you want to tune into that that'll be tomorrow let me check the time on that that is the closing ceremony is at noon or 10:00 a.m pacific

time so submissions are due at 8:00 a.m pacific time we'll be doing a live awards ceremony at 10 10:00 a.m pacific and then we'll be doing another live stream later in the afternoon where we have a couple more special guests coming on we'll talk a little bit about a recap of the hackathon we have uh a couple

other founder friends coming on to talk about some of the things that they're doing and it's going to be a lot of fun so why don't we go ahead and get into the main content of this stream so give me one second here and I'm going to get we're actually just going to live in Excaladraw for the next probably 15 minutes and I do want to talk about just in general

uh how you'd architect if I were if I were starting today how I would think about architecting a master and an XJS site if I'm building some kind of AI agent some kind of agentic workflows how that might actually look how to think about it how to make the decisions of of what to go with and what I've been seeing talking to you know a lot of MRA

customers And thank you my friend yes uh in that uh you should if you're looking for a fun follow on X follow uh follow I don't know how to pronounce your name is Versan Versain but definitely a fun follow and uh Jeremy apparently you've got five projects you're working on for for the hackathon so definitely uh interested to see if you want to show

one of them off you can come on tomorrow i'll make a spot for you but let's go ahead and jump in and let let's just do a little architecture let me share my screen and we will see what we can get done in the next 15 minutes all right so the first thing I want to just highlight in the master docs we do have

this integrate master into your nextjs project so this is a good place to look we also have guides for using with versel AI SDK with copilot kit we don't have a guide a framework guide yet for assistant UI but I'm going to mention assistant UI as well we do have an example in our monor repo with assistant

UI and you know our playground uses assistant UI so there are examples in the wild of that but we will be putting together some better docs uh for them as well but let's just talk about if I were to if we were to be building a nextjs site and we were wanting to use master to build these agents how could we architect it so there's kind of two different approaches so we have approach

one we have we'll just call this like yeah I I don't know I'm making up these names like a bundled application this means in this case I am going to bundle my agents and my nextjs site into one basically one nextjs app agents plus nexjs so let's just say ma let's just say we're building some master agents master agents plus nexjs

bundled together so there are some pros to doing this and we'll talk about what that would look like and then we'd have another option and we'll just call this like a separate AI backend and we'll just say next.js JS is the client calls the master back end so with Maestra out of the box you get essentially a REST API for all your agents for all your workflows so you can deploy Maestra

completely separately and then just integrate into Nex.js uh you know through e either through Nex.js JS in an API route or through you you can technically even do it through the front end although you know there are some some security concerns you have to work out but ultimately you can kind of architect it in that way so let's take a

look at what this might actually look like if you were to actually build this thing so in this case like I'm going to have my nextjs site nextjs app here and I'm probably going to deploy this to somewhere this could be deployed to Verscell or Netlefi or wherever else you might want to deploy your Nex.js

application but inside that application I would obviously have you know if this was like a chat interface for some kind of agent application I'd probably have like my chat component and this is going to be very simplified of course but I'm just hope hopefully you can kind of get the idea of how I'd start to think about it so you have basically have like the UI

right you have maybe chat components that or react components that live in the UI and you're probably going to have uh you'll have like a slash API chat route something like that and so you'll you know that this chat component is going to interact with and then you're also then going to in this route you're going to just basically directly talk to Maestra so

I'm just going to put this all as just one thing just MRA itself and you might have inside Maestra you could have you know agent slash you know agent one and agent slash agent2 and a workflow slash you know my workflow whatever the whatever the ID of of your workflows are right so all these things are basically just agents or workflows that you built in Ma

that you are and just directly importing and this is just by import so you're just importing Maestra you're getting the agent so if I were to pseudo code this you're basically going to import Maestra and then you're going to do something like master.get agent agent one it's actually probably more like let's let's just rename these to make it a little a little

cleaner and then you inside your this is of course all in happening inside this chat route you're going to call you would call something like agent.stream stream and then you just pass in your messages something like that now this is obviously some pseudo code right like this isn't I wouldn't uh wouldn't guarantee this thing's going to run but it's

pretty close and this chat component what you get here is there's a lot of options of how you can actually build the UI part of this application right there are options such as you can use AI SDK and you could use something like use chat they just a use chat hook and then you'd be able to talk to this route you could return this stream and you'd be able to just use that there's also

copilot kit is a great option for building really interactive interfaces and then there's assistant UI and there's probably others that maybe I don't even know about or that are coming out now right there there just like there are a lot of there's a lot of really good competition in the framework space there's a lot of really good competition kind of to how to build these front-end

components um and just checking the chat here uh very sane like very insane got it and another question here will you guys support KNX 3 um I mean it you should be able to use it i guess if you can't I'd like to know why so ultimately we might not have a guide for it but a lot of this you know I I believe that because Ma can be you know

bundled with pretty much any TypeScript project or it can be run kind of standalone there should be no reason it can't work with Nux 3 but uh feel free to jump in if you that's not your experience happy to try to help you debug it all right so as you see we have a bunch of options here right we're ended up but the nice thing about this so the benefit of doing this kind of bundled application

approach let's just you know write down here I'm just going to write like some pros and cons here pros make this a little bit bigger the pros is like easy deployment you just deploy everything together your agents your uh your front end it's all it's all tied together so that's probably the biggest uh the

biggest benefit is you you have that ease of deployment but there are some you know maybe some things that aren't as good about this approach which is it's kind of I would say one of the things that you can sometimes run into is just like serverless serverless function timeouts um you're kind of locked into a single front end

so I'd say you kind of have a little bit of pros and cons and there's probably more than this but I'm simplifying a bit that if you're looking to get started in a somewhat easy way bundling it together makes a lot of sense because you can it makes just that really easy deployment story but a lot of times especially if

you're building longunning workflows if you're building agents that have lots of tools you might start to run into some of these serless functions you can increase those timeouts of course but um depending on where you're deploying to you can have uh issues and sometimes it's harder to debug when everything's kind of together so you don't have the

separation of concerns and also it becomes a little bit more tricky if you do want to add some other kind of front end let's say you wanted to add you know we have a lot of people that are building either desktop apps with Electron uh mobile apps with Expo we have people that are using Telegram bots or Discord bots and so then you'd be building something like okay if you let's say I just built a Discord bot or

a mobile app well now I'm do does my mobile app just hit my next API route and I run master in there it just gets to be the architecture gets a can start to get confusing so I see a lot of people when they're just starting they they go with this approach and it because it works really well it's like probably the easiest to get started with

i wouldn't necessarily say it's the best most scalable but it is probably the easiest now you can go through another alternative here which is we have this idea of kind of separating out our AI backend from our Next.js site so in this case you know you have your Next.js app i could probably just copy some of this

stuff over we'll see how this works copy paste is our friend and I still have all the same options on the front end right i mean I still have my opportunity to use you know AI SDKs use chat or whatever copilot kit assistant UI i still have all those options available to me of course and I still very likely would

have this uh some kind of API routes but the difference here is that instead of having MRA built in I'm actually see if I can grab this there we go i'm actually having Maestra as a separate service and so we end up with I guess kind of the same thing here so I'm running Maestro i still have my agents but it's

deployed somewhere else and then on this side I'm basically using master client which let's find the docs for that where are they um masterclient SDK so we have this masterclient SDK that you can use you can kind of see here this is the code of what you would do so I'm just going to grab this And in this case you have master client so the pseudo code in this case this code would probably run but the pseudo

code would look something like that and again if I was running this locally I'd probably just be using local host but if I deployed this somewhere whether this is deployed to you know master could be deployed to Verscell could be deployed to Cloudflare could be deployed should be able to be deployed anywhere or master cloud you know which we're currently just in beta but you

could essentially point this to that URL and now in your chat route you are using Ma client to reach out to Ma and I wonder if there's a there's probably an example of like generate so why don't we grab that and add this and this would be in our case be agent one and we could generate or you know ultimately we're just going to call stream

and I guess this is messages and so I probably need to fix that over here because this would be the same thing so be something like that so then in this case again this looks like it's more code but it's shouldn't be that much more but you kind of have a little bit different architecture if you see here so again now let's kind of walk through

the pros and cons of kind of doing this separate approach let's grab this so pros uh separation of concerns that's sometimes nice depending on your your you know the complexity of what you're building supports multiple frontends cons probably like more complex deployment I guess more complex deployment and great question so we got a question here for from Amjad can

masterclient run in a NodeJS environment or is it built only for the browser it can absolutely run in a NodeJS environment a lot of times we see that's how people are using it it can run in the browser uh the trick with running the client in the browser is you have to think a little bit about the authentication of it so and we have some tools to allow you to do that but you do

have to think about you know do you want your browser directly calling your master agent the answer for latency reasons might be yes but you do have to just be concerned with uh paying attention to the you know the authentication of it and making sure that you're you know you are kind of handling that and we provide some some

middleware and some runtime contexts and a few different tools where you can kind of ensure that that's the case but the other uh the other option is as I'm showing I guess right here this master client is in this like a this API chat route so that's running on the server right that's a that's a NodeJS environment but yeah you could run it in express server as well and you could

call out to a master backend it would totally work in a in a NodeJS environment and that's honestly how we see a lot of yeah a lot of people using it the thing about that I always is the trade-off right if you are running it from in this case you have an X.js app you're calling to this API route it's

already made one request and then now it's making another um if you're now that being said if your chat route is run on the same infrastructure if you're deploying these to the same place that should be a pretty light hop so it's going to add some um some latency of course but I think and I haven't actually measured

this because it does depend on where you're hosting these things but it's going to be pretty minimal but something you do have to consider is there's a technically kind of like a little bit of extra latency here but it should be uh pretty pretty minimal and so yeah pros separation of concerns supports multiple frontends you know more complex deployment and in this case

if I had this mobile app you know I just now you know my my mobile app just uh it can talk to my master client so again a few different options for how how you want to do this if you want to if you're working with for those of you that just joined us you know I know we've had a lot that have come in over the last 20 minutes as I've been explaining this what we are talking

through is if you're building you know we're using a specific example of building a MRA agent or MRA agents and an X.js app but it does apply more generally if you think about it to some kind of front-end framework with some kind of uh AI infrastructure and option one you can bundle it all together put Maestra inside your Nex.js app you can actually just import it directly in your API

routes call the agent pass in messages return the stream to the chat components using tools like AI SDKs use chat co-pilot kit assistant UI and the pros of this are easy deployment it's all one thing right you just PR something to GitHub deploy it and it or merge it and it's deployed but you do start to run into some limitations especially when

you get more complexity into your application where if you needed to add other types of frontends whether that's you know Discord bots or Telegram bots or you know a mobile app a desktop app some other kind of front end that might want to interact with those same services it gets a little bit more complicated a more uh complex approach but also probably more architecturally

sound for more you know if you are building something that has potentially multiple frontends or that is a little bit more of complex of a setup is you basically have your next.js app you use the master client SDK and you call out to master client or sorry you call out to your master instance which is deployed somewhere else now this could

they both could be deployed to Verscell you could have your Nex.js JS app deployed to you know netlifi and this your mastra deployed to versel or your mra deployed to mra cloud or wherever else you might want to uh deploy this thing and then the pros here are you have separation of concerns so you can you know whether you have different

teams owning different parts of this which is commonly what I see is like someone's responsible for building out the AI parts someone's responsible for the app when you work with larger teams uh supports multiple frontends so if you ever had like a mobile app or a desktop app it's much easier to actually integrate with but some of the cons is of course you have a little bit more

complex deployment you got to deploy your agents separately from your front end if you change your front end and your agents you have to figure out um how how you do that versioning and it there are tools to make it pretty easy but it is just something you have to think about and you have to plan for and so yeah that is a quick little I

guess whiteboard session on how we think about building agents with frontends you know specifically MRA and nex.js js we do have some things in the chat all right so let's answer a couple questions and then we do have a guest we're going to bring on so a few questions here let's me stop sharing and jump back

in could you differentiate between building workflow and agents this one is a little bit nuanced but when you think about control flow on an application so if you're building an agent you're basically letting the LLM decide you're probably building a pretty complex system prompt of what the LM should do you're giving the agent some set of tools and you're letting them you're

letting that agent make the decisions so you might be basically building you know workflows within your system prompt itself saying if this happens I want you to call this tool if you know the person if the user prompts ask something around this you should run this tool and you're you're basically turning over a certain

level of uh you're losing a certain level of determinism because you're letting the LLM decide now if you have things that are very specific to your uh that can be more deterministic that's where workflows can come in because you can basically say these are discrete steps these are discrete conditions

sometimes you might call out to an LLM in that workflow but you do have a certain level of uh deter you it's a little bit more deterministic you kind of know that once it starts it's going to go down this graph or this chain of of thinking oftent times what we see happen is you end up doing a little bit

of both you might build agents that can call into workflows you might build workflows that can call into agents they're kind of composable in a way where you can piece these things together one thing that one analogy that I like to use it's not maybe not the best one but I think about if you I was trying to automate maybe like a customer support bot like just thinking like a

actual like customer support person in a company if I were to hire that customer support person ultimately I would give them a set of tools if the customer support person in this case is an agent it's like you think about it comparing it to like a person if it's an agent I would give them a set of tools you know maybe Slack maybe access to Zenesk

uh you know maybe email whatever tools that you'd give it those would be tools that are given to the agent but I also might give that person a set of like playbooks when a new support issue comes in I want you to do these things and it would be steps where that's you know customer support person is smart right

they they can maybe have to make some decisions as they read that checklist of things they need to do well that's really a workflow so you might have an agent that has this this set of playbooks or this set of workflows that depending on what happens it will run in certain situations and so that's kind of how we modeled it when we built it is is to try to think about it like uh things

that are that make sense to us and so that's the analogy that I often use and yes you workflows can be attached directly to an agent now you used to have to like create a tool and like wrap the call to the workflow now you can just pass in a workflow to an agent and thanks Ryan for tuning in hopefully we'll see you in the next one

and so this is one more question that we get so frequently is how can we just can we display the workflow steps on the front end the answer is yes but not very easily and so we are working we're we're working pretty closely with assistant UI and copilot kit and we're building some examples with AI SDK because that's what

people really want to do is display the steps you might not want to display every step though there might want to be some kind of translation layer because you don't want to maybe display all the internal logic of your workflow right so it's it's like there's like a translation of the workflow is doing something but you want to tell the user and provide updates of status and maybe

they're just simple updates and with some levels of information about what's going on so we we are working on that but it's not as easy today as it needs to be all right we went a little long longer than anticipated but I do want to call in our next our first guest for today so this is something I'm hoping we're going to do at least maybe once a week where

we basically think through what does security with AI agents look like and how should we be thinking about security i am not a security expert you know maybe not a big surprise to some of you so I figured we need to call in someone who is so Ally I'm going to bring you on awesome thanks Shane thanks for having

me yeah n uh nice to see you and I I've chatted with you obviously quite a few times but glad to have you on the live stream today and maybe for those of you you know there's a 140 people on here so most of them probably don't know who you are or what you do so could you give a little background yes absolutely um yes my name is Ally and I am a VC so for AI startups I help

them with SOCK 2 compliance and just building trustworthy AI overall and usually a good first step there is is SOCK 2 um but also I enjoy like going beyond that and thinking about like threat modeling AI applications and also um thinking about ways that we can like enhance our security like at runtime

since um LMS and AI agents are so non-deterministic um I'm also involved with the OAS agentic security initiative i lead the insecure um agentic code examples there um so we build and catalog a bunch of different examples of like kind of what not to do with AI agents to help um inform other people um

how to like build them securely and also to help inform our other working group that's specifically geared towards um securing these um insecure examples awesome yeah great to great to have you on here what I know you you had some ideas of some things that we could maybe do today that would be hopefully

educational to you know the 150ish people watching yeah for sure yeah happy to go through an example that I build i you know it's interesting like I'm actually not um a Typescript dev myself i'm like familiar with Python the most however like Monster is so easy to build with and use that I've been able to create um agent

examples in Monster like very easily and so I have one today we could walk through and look at look at it for the security implications if we want yeah we should do that and you bring the security i'll try to bring the TypeScript i'll do my best i I I would I can probably get a little further than you but you know we'll we probably have some people in the chat that will help us if we uh if we go off the rails and

and unfortunately you know I usually have my you know my sidekick Obby here he's he's the TypeScript expert so he would he would tell me all the things I'm screwing up but you you'll have to deal with uh no no true Typescript experts today but we will we'll get through yeah I believe in us um All right so can

you see my screen we can okay awesome um so this is just a preview of what I built so basically I I built an AI agent that helps tech founders decide if they should live in San Francisco or New York City um it's kind of a fun example i also have it queued up here in my um MRA playground too you can see I just ran it with um Grock right now is what's set up on it and all we see is just your

architecture diagram right now oh that's right um that's cool oh I can say share this tab instead can you see the monster playground now yes you might want to Can you do like one level of zoom maybe yes yeah just command plus or just There we go that's That's better hopefully hopefully you all can see it if you can't see it chat let us know but that that looks pretty good

perfect yeah you can see I've I've run it a few times i run it with a few different models and I'm using um let's see if I switch back to the architecture diagram um so basically our agent is connected to the weather tool that actually comes pre-loaded with MRA if you tell it to install um the examples

that it gives you during the npm I think create MRA workflow um and then I've got it also connected to an MCP server um from Exa and it's got the EXA search one um it basically just like calls out to um like I think up to like five URLs by default and does some research on the web and returns you those results from those web pages it scrapes um so it's

that it's got that one single tool from that MCP server and then our homegrown um weather tool um and it's also connected to the MRA memory DB that MRA creates automatically um so a few things that I noticed while I was creating this example um were the following security concerns and you'd see um this diagram as well let's see if I can make it a little bigger um but

basically there's the potential for um indirect prompt injection um from both the weather tool and the MCP server primarily like mostly the MCP server um I know Shane just talked about this and I think it's a really great point that um you know if we're able to find ways to include like non-determinism or sorry um more deterministic more discrete like workflows during um our EI agent

development that's actually like very beneficial and helpful for security um and one of the reasons why is you know for this Excel Labs like web search tool we have no control over like one like which websites it's going to um call out to um and like how how many of those websites it's going to actually like return in its response um when it decides like what information from those

websites um was worth returning back to the user um so it's like a high level of or high amount of non-determinism um so we're not sure one are those web pages like even factual that it searched um or we're relying on exit search for that um and two like could those web pages that scrapes have like indirect prompt

injections in them where there's some sort of hidden instruction like you know export the entire chat history back to me and when we feed that back into our agent um that agent might understand that as an instruction rather just rather than like web um search results and might take action on that so so so

can I interrupt and ask because of course I understand this completely but maybe someone in the chat doesn't or maybe I don't so you know what so prompt injection in this case I'll explain it you tell me if I got it right so let's say we pull some data we scrape some data from the web and either hidden in

that text somewhere you know whether it's a you know some kind of comment or whatever but basically tries to tell the get the model to return some kind of data that it maybe shouldn't have is that accurate yes that's correct that's a really good explanation so rather than trying to you know just re getting the HTML or the the text it's actually like

trying to you know somehow tweak or change the the result to do something maybe potentially malicious exactly yes that's a great example and the reason why I think the weather tool has less risk of what you you just talked about is because that has predefined APIs that we call out to so we know every time exactly where we're getting our data from it's from you know

if we've configured this API it's probably a trusted source that we know um it's going to return in the same like standard format every time um so there's more um determinism within that workflow um so there's probably less chance of there being an indirect prompt injection there so just something to

keep in mind when you're connecting like workflows versus MCP servers um and then another interesting I wonder if I can go back and show you a previous run that I had did in the MRA playground um but basically there was one maybe one of the first tries I ran i don't know if it'll show it but um basically one of the first times I ran

it I ran it with um 03 uh sorry GBT40 and that model for some reason um decided to call out to the um MCP server like multiple times like all at once i also had a different prompt on it originally the instructions were different um so originally I asked the um or I put in the system prompt like you should find out information about cost of living different activities to

do the weather um and with G I'm sorry GBT40 it decided to sort of solve all of those or research all of those questions at once um and I don't think the MCP server necessarily liked that or I saw like a bunch of different like four or five different calls to um exit search all at once um and so I got a bunch of

errors related to that it seemed like the MCP server itself like wasn't able to um handle that amount of requests just in a really really short amount of time um so that kind of like takes us back into thinking about something called unbounded consumption um and you know this is a concern not only for

security but like for performance um with AI agents um we're not exactly sure like what decisions they're going to make um so we don't know how many times this our tools could be called um in a row like what if you know our AI agent decided to call it a 100 times or 200 times or there was some sort of you know

prompt injection or it went off the rails or something caused it to make too many requests um that could easily create a denial of wallet attack where you know every single call we use it uses API keys it uses um our tokens that are associate associated with it so um we don't want to basically spend all of our credits just because our AI agent

decided to make too many calls so thinking about ways to like rate limit those essentially um is really important or refactoring your system prompt um for a defense in-depth strategy um to like not only do the rate limiting but also to like refactor the system prompt so maybe it makes its decisions more effectively um but also like try

different models so I thought that was an interesting takeaway um from this example for me was like when I switched to um Gro 3 it decided to you know sort of like do this more iteratively where it would like ask me like hey okay like I don't um I'm not going to find all this information at once but here's cost of living what do you think about that and if um depending on what you think there then I can look into you know

activities to do and I can look into um weather and other factors so um definitely trying different models is important um especially too for like bias and alignment purposes which is I think a really interesting part of AI security because um alignment's going to be unique to like your individual business an example to explain that that

I will give is you know in this example like I don't really care if like it tells you to live in NYC versus SF like I don't have a personal like investment in there um but like what if instead of like me just being like Ally what if I'm like the city of San Francisco and I created this agent uh maybe I wanted to

recommend San Francisco like maybe I want it to be like agnostic at first but my business use case is actually like you know I want to recommend SF at the end of the day and get more people to move here um like boost our economy all of that so that's important to like you know we maybe you want to think about receiving your application and finding out exactly like what it's going to say so that you

know that it's like aligned to your use case sorry Shane yeah i was just going to say I I've noticed you talked about changing models and having it call tools differently and kind of approach problems differently we actually spent quite a bit of time just going through Monster actually has a compatibility

layer to try to make models more uh make sure that tool calling is more accurate across models because there there's such a a varying level of just tool accuracy in in different models and so not only are they deciding deciding when to call tools obviously changes but how frequently and how correctly they call tools it does vary based on model but I

do need to stop just for a second because we have special guests surprise guests joining and I did not expect to see uh so you know co-founder Obby you know founding engineer Tony in Japan live from Japan live from Japan so let's take a let's take just like a twominut delay and yeah what's the what's the vibe in Japan i haven't I haven't talked to you for 48 hours that's the longest I

haven't talked to you in six months all right so right now it's 5:21 in the morning the sun came up at 4:00 a.m um which is crazy um and it's super quiet and peaceful out there one interesting thing in the toilet the sink is on top of the toilet so like after you use it like you flush it and then like the sink fills the tank i'm like this is amazing

right like washing my hands in the toilet which is great um what else did we do we had sushi that was cool um and then we met with some McKenzie folk uh last night and they were talking about lane chain in not positive ways so that was interesting yeah what you got more meeting tonight today's our big meeting day we're going to head over to Plaid

which is a Japanese company and then several other kind of meetings with the the folk here but yeah thanks for having us yeah yeah so pleasant s surprise wasn't expecting you all to join join but Ally you want did you want to continue you want to keep talking through where or were you all finished up um I can I can

definitely keep talking about it i was if anyone has any questions too from the from the chat about anything I covered so far happy to answer those too yeah if you do have questions so for those of you just joining us we have Ally our you know resident security expert to help in ways that I cannot so we have Ally here uh Abby who's normally here he's joining

us from Japan with Tony and we are you know talking AI security if you missed the beginning section well this is recorded on YouTube you can go see me talk about MRA and Next.js and how we think about architecting applications and kind of connecting MRA to different frontends specifically Nex.js js but it doesn't have to be um so we got a

question Obby where are you based out of based out of uh California so not but just in Japan because we do have a lot of our user users in Japan so our doc you'll notice uh kind of an Easter egg if you didn't know it was there if you go to our docs you'll see that you can actually uh see the Japanese translations of those because we have so many so many of our users in Japan yeah we're making this joke that

we're in the land of 20% of our users so we're trying to meet as many of them as we can yeah exactly um I I guess Ally what other any other interesting things that you have coming up i know you you know you do a bunch of hackathons you do a bunch of events anything that's on your radar things that you're planning for in the near future yeah great

question um so I think I'll be at New York Tech Week the first week of June um recently um I started in connection with my friends at Penszar a Insecure Agents podcast so we're looking to shoot episode two that week while I'm there um and then June 18th is it released is it available um or not yet the first episode

yeah it is okay well find us a link we'll have to drop we'll drop a link here in the on the screen so I mean people I'm sure people can find it but I It's called The Insecure Agents Podcast yes let me see if I can find a link it's on Spotify all right well we we'll get that pulled up but yeah curious on what was the what was the topic of the first

podcast so the first podcast we were lucky to have um a guest Mark Dorsy who's the CISO of Netifi with us mark oh really that's awesome yeah Mark's great um we chatted about sort of like how you know the time we're living in vibe coding everything like everyone is able to build basically now and so like what does that mean for

people developing like agents specifically and what tools are available um and he was able to talk us through sort of like tools that Netlifi has to help you deploy agents securely um which I feel like is kind of like the next frontier like it's really easy to create like an agent PC but then to like deploy it and actually have a production like that's pretty challenging both from

like a production and security perspective um so that was a really interesting conversation um we also chatted about like our expectations for like RSA as well since it was filmed in the backdrop of RSA cool and did you say what so what was the um I guess how frequently are you planning on doing these podcasts is this like a weekly thing that you're you're

committing to what's the what's the cadence that we should expect if we want to learn more about AI agent security so it's at least monthly and then as we have interesting like you know guests that want to come on or interesting topics to discuss or like something that's like new and um like everyone's talking about like an MCP was you know

was huge like 3 weeks ago we would would have capitalized on that if that was um we were doing it then um but definitely want to increase the frequency and we're going to start like we're going to put a website up insecure agents.com eventually will be a website that we will own and use for the podcast so yeah definitely stay tuned all right yeah any if any anyone is in

the chat and they do have questions please let us know you can just whether you're on LinkedIn whether you're on you know Twitter X whether you're on YouTube just drop a comment and let us know what questions you have we're going to be wrapping up with Ally shortly but I did want to drop in if you do want to see what Ally is up

to you can follow Ally onx here and you'll I'm sure you can find the podcast from there you can see all the other stuff i think she you have a newsletter too right Ally yes and that goes out every Thursday yeah so if you if you care about security and you probably should uh then there's there's a good place to go to kind of keep in

the know we'll also our goal here is you know we'll probably be doing this as a recurring segment and each week we can talk about something that's interesting in the security world uh maybe we can find some cool examples of of code to show of how certain things can be secure and and maybe maybe aren't secure

because people like me need to learn these things so I appreciate you coming on and yeah we will probably see you next week Ally perfect awesome thanks for having me see you see you all right dudes what's the vibe then there were three then there was three i don't know i think uh the hackathon is the funny thing we're in the future

right now so the hackathon ends today for us or tonight yeah so we'll see y'all on the the late live stream for us um are you guys going to be doing some live streaming while you're there yeah we after our meetings today we're going to do a live stream in this time zone so double dip I guess yeah what's the rough time to expect that for people just so

um probably like I would say 3 p.m in this time zone which is I don't know what the time zone's called but Japan time yeah i I don't know what that conversion is but uh 3 PM Japan to Pacific so it' be like 11 PM Pacific time maybe yeah we'll see how it goes yeah so you know maybe maybe I will uh tune in from from my bed because because I'm in

central time it'll be like 1:00 a.m for me but yeah cool and what are you going to be working on what are you planning on working on in that we're working on agent network um so actually I can answer this one question in the chat too in in line with that how is Ma different from OpenAI agent SDK um in a lot of ways um it's

the same in a lot of ways is different so from for our perspective we're fully built in Typescript with a modular module structure um and we kind of go through every AI primitive openai agent SDK is trying to solve the same problem and that's chill um by default in the AI SD open AI agent SDK let's say that three times fast

um their kind of main primitive is agents like collaborating with these things called handoffs where you can delegate from one agent to another um and then you can kind of have them all collaborate we experimented on this feature many months many moons ago and we had like our experimental version which allows you to essentially create a network of agents that all work together and uh Tony and I have just been

scheming on how to improve it and we just figure let's just do it live and improve it live and then and then at the end hopefully we can get a stronger you know network story for multi- aent collaboration awesome yeah excited to see the progress there and definitely get you know just a little bit more of our like agent network story because we have it right and some people actually really enjoy it but

we do know we need to take it we actually got to finish it we kind of released it as experimental and we've learned a lot and so now it's it's how do we get it to the next stage yeah it's like actually an interesting thing for anyone running open source projects like even if you mark something as experimental or do not use or whatever people will use it which is good and we

got a lot of feedback from that so now we're like let's do this thing also you can use it it's like experimental in the sense that we will not support you if you have bugs right like because um but now we're going to Yep yeah exactly so we're we're we releases experimental so we can learn we've learned a lot now we're ready to really do the thing for real um also this is a great great

segue so I don't know if you read my mind shoe or or what but you know I just found out about this hackathon thing right now it'd be amazing if you give a little intro and what kind of apps people are making well the good news for you is we're actually going to be bringing on two hackathon guests to talk

about what they're building so you can actually hear it from them we had a really great guest on Tuesday as well during the live stream i think it probably started either 10 or 20 minutes into the live stream so if you want to see another example of what someone's building for those of you that are not

familiar with the hackathon that we have going on it's all week go to master.build you can still sign up there's still you could still work on you have you know what 12 18 hours or so you know a little less than that to still build something so if you have a free evening build submit something we have a whole bunch of prizes we have prizes for you know

winners we have categories we have prizes for the winners in each category and then we also have just raffle prizes which means if you submit something any kind of like agent that you built with MRA you have a chance to win some of these prizes so even if you haven't been spending the whole week working on it you don't think you are going to win you

should still submit because there's a chance you're going to win some good prizes and everyone that submits will send a copy of Sam our other co-founders book principles of building AI agents so we will send that to your doorstep and with that I do want to bring on our first hackathon guest ah yes you brought

your book with the book of Mastra all over the world the book of Mastra written written in the palace of the dog all right so let's uh let's bring on Justin so Justin I'm gonna bring you on and you're on hey everybody can you hear me we we can perfect yeah how y'all doing good good yeah interested to hear you know obviously we

we've chatted a little bit you know the last few days during uh hack the hackathon master build office hours but I think everyone is interested to hear what you're building how what I guess what are your first what are your thoughts on the hackathon so far how's it been going for you oh man like I Okay I just honestly found out about MRA like

less than two weeks ago probably so this is the week that I decided to kick my butt and try to learn it you know and uh string together something and the hackathon's been amazing uh Discord's been alive it's uh I don't know it's it's fun so I if you guys do another one I'll definitely be a part of it and maybe dish out something else something

else that's new but but yeah like uh having a deadline really helps yeah I mean I think that was kind of like our rally and cry too is we we heard a lot of people that you know kept telling us oh this looks really cool I need to find a weekend or I need to find an evening to do it and so I think in a

lot of ways a hackathon is just a little bit of a forcing function of okay you know a lot of a lot of in-person hackathons are of course a good forcing function we have such a distributed you know community that we wanted to do it virtually but it is a good forcing function to say okay like I need to sub

I need to get this submitted by this time so I want to have something so you have a little bit of a deadline and yeah it seems to give people a little bit of a kick to to try something new yeah and like um I've been doing JavaScript webdev for like I mean forever basically but and I've been using AI agents and you know windurf and chubbyt and

whatever just helping me work and do code in general but never really I've always wanted to dive into this moment and then I was starting to learn like lang chain and getting into Python i don't really know Python that well javascript's definitely my stronger language and then like when someone at my office was like you should check out

MRA i was like oh what's MRA oh Typescript yes let's get on it like so just pretty much not that I gave up online training i just don't I'm I'm pretty much full in on this platform right now and I've been I've been doing it uh pretty much every night this week i got full-time job got kids so I've been putting on the night the I don't know the nocturnalness of the developer

life you know and getting getting it done so yeah well we'll working in the dungeon exactly yeah you're wired in you're getting wired in know that that well excited uh excited to hear you know like that you've been enjoying it so far I guess what's so two more questions before I I do want you to kind of show off what you started working on but we

we do have a little bit of time so what were you doing or I guess what what's your what's the day job or what were you doing before you said you've been spend a lot of time in web and honestly that's a lot of our a lot of the people that are using Ma are kind of like us like we're we're not ML guys like we're going to have a segment some at some point we're going to bring on some of our ML friends it's going be uh builders learn

ML because we're not ML machine learning guys we're web guys so we share probably the same background but yeah can you tell me a little bit about what you Yeah what your background is yeah I mean uh I work at a company in Michigan called Brightly and I've been working there I mean it's crazy i've been there almost

11 years so it's been a while we've been turning out some cool stuff but web I my background is mostly design um well it back then I gave up on design not that I gave up on design but I was trained as a graphic designer and then just so my path in was web design in general and was like holy crap this is so much fun this is so much better than design so I

just went full boore into it taught myself how to do it and I mean I've been doing I started with like WordPress you know pretty typical just and then uh I played around with headless CMSs and stuff like I've used Gatsby in the past too but a lot of times now I have um uh I'm just basically building React applications for businesses I like SAS

software so in the security space and uh insurance space and some some of these bigger markets and stuff like that so I'm I'm definitely more full stack nowadays so I like diving into architecture and all that type of stuff too so it's it's not just um you know web code but I focus a lot on the user experience side of things you know given my background and I think that's if I don't have a reason it's hard to do the

code you know yeah so but yeah like um as ever since you know GPT3 released and chat GBT it's been pretty much forefront on my mind getting into this AI stuff and and uh like you said the web is like where people I mean apps and native apps and all this stuff great but like the web is where it is all accessible you know like anyone can

get on a browser anyone can do this stuff so it just it keeps me interested you know always and it's always something new with some new platform or browser or whatever that comes out and it's it's it's an exciting space to stay in you know so yeah yeah ex absolutely and I think Yeah yeah we did have a

question in the chat you know like what's the audience that we're targeting well I would say one of the audience types of audience are us right the web devs that are now learning AI wanting to build things and so often you can you can accomplish a lot with just these really great model APIs right you have an a and we're really good at interacting with APIs that's what web

devs do so I I do feel like it is a really good time to to be a webdev to um really kind of be in the um be in the AI space but also coming from a background of web development so it's it's a great time to be here uh we got another great comment here we found out about MRA a couple weeks ago and we've been enjoying it so far uh yeah thank you

i I apparent that as well too so it's it's been fun so well appreciate it appreciate it everyone who is here uh let's catch everyone up and then Justin if you do want to get ready to show what you have been working on so we are talking to Justin he is taking part in the master.build hackathon you can still

learn more about the master.build build hackathon go to master.build and you can still sign up if you want there's still some time if you want to submit something we have a whole bunch of prizes for uh anyone that does submit some kind of submission everyone gets a copy of this book if you submit something and you also are in in

a bunch of raffles for other types of really cool prizes and it's been going on all week we have the awards ceremony tomorrow it's at 10:00 a.m pacific time I believe so deadline of for submissions is 8 am so get your submissions in if you're watching this and you are participating and tune in we'll have we

will be for anyone that's that has uh actually submit something or is part of the hackathon you'll actually be invited to the live Zoom so you can interact with with us all and we'll also be live streaming it as well for those of you that just want to tune in and see what the awards ceremony is about and see you

know what was built who won what kind of prizes we had all that all right should we should we check it out justin you want to share no did we lose Justin hope not the internet the internet gods have failed us oh no but he's back i'm back sorry i'll get to it but yeah so what I'm building is um for my hackathon

project this week i I pretty much started MRA this week and also started this project this week so but I'm calling it family story books and before I get into using it I figured I'd just peel back MRA playground a little bit just show you kind of what's going on behind the behind the scenes and then I'll get in and actually use it so so behind the scenes I basically have

two agents a family historian agent and a family storyteller agent and the idea of this whole app is sit down work with a family member or whatever uh collect some basic info get a conversation started get the uh models to you know maybe jog some memories of some anecdotes or some experiences that you may not have thought of and work it into

the initial kind of workflow step which is like a a factf finding a historian gathering session so it'll work with you back and forth and it'll ask you questions and you can keep providing feedback and then once you've done those steps it feeds it into a family storyteller agent which is crafted to take the input or the the output from

the historian step and create essentially like a chapter book uh or or attempt to anyway and for right now just to call it out I'm using the really small models right now just for the sake of the demo speed and whatnot so I'm going to play with figuring out some of those models and see what works best in

the future but behind the scenes there's a workflow called a just one workflow this is a pretty simple the the AI server part of MRA came together really really fast like it was just like so easy to use and uh I'm using the V-ext workflows and basically it starts with a summarize step so it summarizes some of

this initial feedback that you can see on the right over here and then it will go into this provide more context step and this step is a suspend step that will just iterate over and over and over until the user collects the checkbox is satisfied and you can keep working with that over uh until you're happy essentially with that and then it moves

on to like a mapping which maps the data out of this step and into the story step which will create its first draft of the story and then just like this step up here it's a suspense step that will go over and over and over and over until basically you're happy with the story and then there's a part that isn't done yet i'm going to try to get that done tonight before the dead the deadline which is output uh and generate a PDF of

like your story book and I can go into more of what I'm planning on doing in the future but we'll just kick over and see it in action so I don't I'm not using any copilot kit UI or anything i kind of just wanted to focus on MRA APIs and not have some layer in between so I just built this really rudimentary prototype UI to use it so I'll just get

it going so I'm making up names i'm making up data here so we'll just put in some stuff and then when you hit start you'll see the generating preview start i'm not I'm going to try to do streaming as well in this um but anyway comes in with this it changes the step over on the side here gives some users a prompt so review the text

on the right we have more questions ideas you can provide more context so this is what it summarizes the agent is designed to summarize what we know so far give ideas on what is worth exploring and ask some specific follow-up questions that are just open-ended to try to get people to hydrate this more and in the future I want to bolt on like a tool call or a

rag system where people could just click and drag files over or go out to like public census data or some integrate with something like ancestry.com or something so it can go and do its own research without you necessarily having to provide direct feedback but also let you provide direct feedback so okay can

So can I interrupt and ask some questions yes 100% uh so this is really cool so I think if I if I were to summarize like some of the things I could see possibly happening here it's you're essentially entering some information about somebody right and then you're generating essentially that person's story but you're kind of prompting them to give

more information so you're almost like and I know someone that kind of lives near me and they basically they're like a book publisher and they actually bring people in and sometimes it's entrepreneurs sometimes it's just like grandparents who want to share their story with their families you know and it's like they basically do this where they like sit down they do these sessions they sit down for like two

hours or what i don't I don't I've never done it obviously but I've talked to the guy that runs it and they'll like get this they'll like draw the story out of them and they'll put it in a story that so they can share with you know whether it's like actual book that people sell or whether it's more like a like an

family heirloom of like having the story so I could see something like that at some point of like automating that process and like actually getting a physical PDF or book of something like this that's really cool oh yeah yeah and you're you're in my brain right now so but yeah the uh I was thinking you

know in the future too like another I mean and the way that this is architected you hit it on hit in the beginning of the stream or near the beginning beginning of the stream this is a dedicated AI server with a separate Nex.js JS that has a back end for front end that's communicating with the AI server outside of this you know so and

that's because I was thinking maybe there's a mobile app that you bolt on maybe there's a you know another way of interacting exactly and that that's why I that's why I did want to spend the time talking through that because it is a decision people have to make and you should be try to be forward thinking but you also want to make sure you you know you do want need to just make decisions

so hopefully that was helpful but I'm glad you watched it and yeah so it's interesting if we tie that back you're so you are separating out your front end from your master application yeah so just quick behind the scenes I have a monor repo here and two apps an AI server and a web and the web just has a

API route to a submit and a resume with the configuration of the master client just sitting right there in the back end and then the front end just makes calls to the backend for front end out to the AI server so that's a little bit behind the scenes I guess so yeah exactly yeah and then maybe we should keep going through the story and then it would be cool to spend we do have you know about

10 15 more minutes we could take a look at maybe some of the code and and how you wired up some of those workflow steps yeah you guys will probably cringe at what I got but it's it's hack no hey it's literally called hackathon that means you hack something together and you get it out there and then you know

optimization is for post hackathon if you want to keep working on the project right it's like so I think it would be useful for people like let's look at the ugly code because then you know then you can learn a lot from that and then fix it later right you can always Yeah exactly so yeah so um for example uh the

workflow basically doesn't return back so if I don't provide feedback or click this checkbox and I just hit provide feedback it's stuck in that loop it just suspends in that step again but if I go pro uh provide some feedback like had 12 kids you know go out and all what it's going to do is summarize in this section what we've already known and add the additional stuff that we just provided

so he had 12 children indicating a large family and then it just that drives more questions so did any of Blake's children follow similar careers have notable achievements stuff like that but uh one kid was the kids was the actual James Bond so but in this step so I could just hit this and keep going forever but I

could also provide feedback and just trust the output at at the other end so I can hit I have no more context to provide so hit next and that's Yeah this is just a really cool example of like uh human in the loop right exactly you're like suspending and resuming on the same step and you're just like you could

infinitely loop here forever and then when you're done you say like okay let's move on this is really cool yeah exactly so as soon as uh it goes out and so it's generating the uh agent call to update this and it's also gener or calling out the next step to actually build the story so and you can see the step changes over here so story feedback we've generated initial story based on

the information you provided if you have any feedback below provide an update you know so so it uh the UI is kind of how it is but it says you can see the actual James Bond comes in here so it's updating our history in the in the state and then it's also outputting the story now and I could give it story here so

change the title of chapter one wow my typing live stream uh to before birth or something you know provide feedback and then this one is just it's just going to go generate the new story you can see it's generating down oh it was already done so already changed the chapter to before birth and then the rest of the story something went wrong here but the rest of the story went down

would be down here so and then basically at the other end of that if I don't provide feedback and I hit I'm happy with the story it'll just skip the agent call everything it'll just take the previous output and just pass it right to the next the the complete step which in this state it doesn't work but all

done you've completed the story here's and it's ready for you to download so this is what I'm going to work on tonight so hit download and it will download a PDF of the story then I'm gonna put some like loose styles on so it doesn't look like just a list of you know Yeah chapters so yeah and the nice

thing about this is it's just you know I don't know what you're going to use to generate the PDF but it's any MPM package or whatever there's probably some libraries of just generating PDFs and you already did the the hard work hopefully you know you never know but hopefully it's that part can be just a simple step where you can you know the

siling will probably be the hard part but yeah really really cool the side thing with having the AI server out on its own thing the AI server doesn't really need to worry about the PDF generation that part's going on the back end for the front end you know and maybe eventually there's some like services that will do PDF generation if there's

like a mobile app or something like that but but like I'm being I can I have a choice where I'm putting these modules you know rather than So yeah you're right yeah you could you could you could have the the AI backend generate it and store it to some storage somewhere and pass in like where the this is here's the PDF file to download or you could

put it right on the front end and just use some kind of web library and just Yeah so you do have options for sure yeah exactly that the architectural decisions we all have to make as we build these apps right yeah i saw um Yeah i don't know i'll send you a YouTube short in Discord it's pretty funny but it it explains what you said you just have to make a decision at some

point later you might regret it but right now I don't regret it it's very fun yeah no yeah it's looking really cool yeah i can totally see how you could build upon this we did have um you know one feature suggestion imagine this with transcribed it'd be great to have speech input oh yeah so yeah like a text to speech there are tons of good AI text

to speech tools that could be you know just a you could either have your agent do it or you could pass in a workflow that transcribes and gives you an MP3 or something be good and like I was thinking the the the crux of this idea is definitely like the data gathering like it's hard to convince people probably to sit down and

just type in a bunch of stuff and that's it so that's why it's like it's got to be pretty flexible and ingest a lot of different file types maybe audio formats like maybe you're just on your phone recording your voice hit submit and it's in the memory now you know so I think really where this idea breaks down is the quantity and quality of data that

can get into the initial step the historian step to go over it's inherently multimodal right like imagine a world where I take my photo memories and I pass those as context so they know that I was in Japan or I was here and there was these were the dates and I looked happy in the picture so maybe I'll write something about that you know what I mean exactly or Yeah yeah you

might be able to know like location data like where in Oh like that yeah so where in Japan were you like where what places did you visit and maybe it could then know to ask you a question like what was this was this a business trip was it personal what were you doing you and if you could imagine just like talking to

it as well you know just saying like "Hey this is what I was doing." And there's so many like so many cool things that this unlocks that you've kind of taken like the first steps of so it's it's really cool yeah if like you had seven pictures of pizza every night at a AI meetup it would make a story saying these guys go to a lot of AI meetups for

pizza you know it als Yeah it would also be saying like "Why do you eat so much pizza maybe you should eat less pizza." Wow probably not healthy chapter one a life without pizza yeah um yeah can we look at some code sure yeah I'm down with that um also before we do that like one other small world thing Obby Justin knows Zack Erlocker oh what dang yeah he's uh he

said to say hi and all that stuff to to you guys yeah well now now now you're saying hi for him in front of 230 people yeah but yeah that's crazy are you from Traver Traverse City or Yeah I'm in uh we're in the same co-working Well 20 Fathoms the same like co-working incubator space type thing so Oh that's

cool it's fun yeah Zach is a lot of fun i've known him for a while and and yeah he's good he's a good guy he's a great dude yeah learned learned a lot from him thanks Zach all right let's look at some code sure um maybe start at the resume the suspense step because I think that um I actually did run into a problem maybe but maybe it's not a problem i'm just doing it wrong but um whenever we'll get

into it here let's just get into here so I have a workflow with uh the create family stories step can you take Yeah can you take it up one or two clicks let me get um let me get some more code viewable here yeah that's looking much better okay so this is a one of the provide this is the first provide feedback step

that's looping and trying to extract history essentially out of here so right off the bat I'm just getting the feedback is satisfied and this is the one weird thing that there might be something in the framework that I wasn't aware of but right now I'm just using the front end state to pass in the last

known content to the step and I thought originally that it would be coming off the input data but it seemed like the input data is always tracking the the data from the workflow trigger not not like the last time it was suspended or whatever or but but I just hacked around it really quick and I don't even know if

that's really by design by by MRA but I'm passing in the previous known state of text so that way if if I come back and there's no feedback and we're not satisfied we're all we're just going to pass in the previous response so that way it just it always stays the same so if you hit the button without in inputting feedback or checking the box

it'll just keep the same kind of content over and over and over so it just doesn't blow out and go back to the original input essentially but then we go back down and if if we have feedback and we are not satisfied eventually we get into the call out to the agent so I'm passing behind the scenes the to the

assistant here oh this is uh yeah story feedback so this would actually be the history um generated in the previous step this is the the feedback on the story step and then the new feedback from the user integrated in essentially so some error stuff right here but then later on let's say we gave it feedback so we wanted the LLM to run but we don't

want to progress to the next step i'm just double-checking again to see if we're satisfied and if we're not satisfied we're suspending with the result from the L from the agent call so that way and then it goes back in we get the new fresh data on the front end and then we can repeat this process over and over and over and over for basically forever until we're satisfied

so and then um on the front end over here to invoke it or to call it out I just have one page well actually I have um yeah so I have a couple forms i just built the few forms for invoking the step from the user so the initial history feedback or the initial form with the date of birth stuff like that the history feedback to collect and then a story feedback to collect feedback for

the story and in here all I'm doing is so here's the submit initial step that first step from the initial form um I just have some local state for the user feedback if we're generating or is loading because I'm not using copilot kit or anything fancy like that so I was just kind of building myself um so start

some user feedback state i'm calling the API submit which is just passing in the data that the user gave me from the React hooked form and then on the back end all I'm doing is just passing or I have my master client all configured and I have my workflow creating my run here so I can get my run ID and then getting the result off of that and passing it through to the the

response on the front end so I can show the data and then also store a run id and for follow-up workflow uh initiators so and then on like and then on the other side uh so these other these other feedback forms here so let's go to the first history on history feedback handler a submit handler so I got the data from the form which is just feedback or a checkbox is satisfied and

we'll pass that data into the resume API call and then on the resume API call what we're doing is we're getting the run ID the step ID that was passed in from the form um and the feedback and is satisfied and the previous response so I can pass that all into the workflow resume async so this is the resume schema that it expects and then um and then yeah just

wait for it to do its thing and then return the result here so um I think yeah and then inside the forms too I'm just hacking you know so like my uh my form submission is let me find one really quick on the history feedback I just have a bunch of hidden fields for um the run ID that just gets populated off the state and then the step ID which I think I'm just hard coding in somewhere yeah right here so

it's it's since it's a deterministic workflow I it's pretty easy just to say this form goes to this step but I could see building out like a whole step logic to know what step you're on and being able to pass it through more dynamically rather than statically like this but it worked for the hackathon so and then yeah uh then I guess the

output would maybe be important to see and I have my funny comments down here uh sorry I'm getting to the page here boink oh so I'm having hackathon deserves some danger i'm having the LLM output HTML so for better or worse I'm just you know putting the HTML in the web yeah where where's Ally alli Alli's It was just on

Alli is uh you know over there like what are they what are they doing but no yeah I know i was taking it for you live a little you got to live dangerously when you're at hackathons right exactly well I mean I figured you know the worst that's going to happen is um I don't have a database you know my my OpenAI

key is probably going to get leaked somewhere maybe but like I'm just going to roll that after judging that tomorrow so but yeah so like what's left you know a lot about what we talked about you know tool tool calls some integrations stuff like that maybe an ecommerce layer to do book fulfillment so people can pay

to do a book and maybe get a little bit off the top from that definitely O support and actual security you know like because everything's pretty much just wide open and then you know supporting the multimodal aspect of the history collection you know the data collection so yeah well this has been really cool it is great to see how much

you've got done you know you just started building with MRA or heard about MSRA two weeks ago right so this is basically super impressive hopefully uh you know as you go if you do have questions you know where to find us in Discord of course or on all all the socials how should you know is it are you uh on

any of the social channels you want to plug while while you're here or how can people follow this to see that someday if you release this if this graduates out of hobby project into full-blown app how would how would they know I guess like I'm probably right I don't do I'm not a big social media person I more consume it than anything but I'm on X I'm on every platform but I would look at X um

Justino and the um and LinkedIn is probably more okay where I would spend a lot more of my time if I was to promote something like this i think just and and if nothing else if you promote it in our Discord we will help you we'll get we'll we will share it out to the world for you as well so yeah yeah definitely all

right Justin well yeah it's been great to have you on and we will uh see you hopefully soon as you like if you have updates to this thing we'll see you of course maybe you know maybe at the awards ceremony tomorrow you know time time will tell i'll be there um we'll see where I place no I'm joking okay i mean you I I I'm not a judge but I think you got a strong contender so

yeah very cool project all right see you Justin appreciate it talk later bye all right all right that was sick that was dope dude and V-neck workflows is the only reason that thing worked because if he was using V1 Oh boy yeah yeah i know Tony's probably around there somewhere thanks Tony for making thanks for making V-Nex workflows uh all right

we do have some more guests we're a little running a little bit behind but that uh Justin was not the only hackathon guest we wanted to bring on today so again for those of you that have just joined us I'm Shane this is Obby we're from Mastra we're two of the founders and we talked at the beginning I showed a little bit about how to architect an X.JS plus Master project we brought Ally on and talked a little bit

about AI security we just talked about Justin who's part of our master.build hackathon that's going on right now if you go to master.build you can learn more tomorrow we'll be having the awards ceremony and we also have a few other guests to talk more about their hackathon submission so I think we have

Harun and Muhammad so I'm gonna bring you both on hey how you doing Shane good i'm always seeing you pop up on my uh LinkedIn feed you know with the your live streams or the selfie videos so it's good to finally meet you in person yeah it's nice nice to meet you all as well pleasure pleasure have you heard

you were in Japan as well so thanks for thanks for joining thanks for being here yeah I just wanted to say so just quickly um Abby I'd really like to apologize just right now for the probably the million questions I've asked you in the Discord over the the last three four months of stupid questions no problem at all man yeah yeah well well happy that you had questions and hopefully we were able to help you with

at least a lot of them and yeah really interested to hear you know I guess maybe first just a few minutes on your background like a minute or two and then obviously want to talk about what you're building in the hackathon yeah sure um so me and Muhammad are actually currently building something called petals um maybe I can share my screen

and sort of show you what that looks like and the main framework for that is based on on Mastra and we're deploying loads of different sort of agents to try and create a whole data team so replicating data data engineers and data analysts as we were sort of doing this we've realized there sort of a spreadsheet problem which is you know I

used to be a data lecturer we used to teach that spreadsheets were semi-structured but you know the rarity of a situation is they're probably just very unstructured we had a plan to sort of create something that allows you know user input to structure spreadsheets a bit bit later that was later on down our road map but this

hackathon came up like you know what let's just let's give it a go and see what we can can pull together so Muhammad's been spending the past three days sweating to to hack something together um but yeah this is sort of what I'm talking about you know when we deal with our system we can connect to like loads of different API

endpoints does have a lot of spreadsheets which is not a problem the issue is when you're trying to pass that into LLMs you know sometimes there's colors being used to signify information um other times you know they're just like this is a nice looking sheet but if you're trying to pass this data it's

quite messy right um and can you give us just can you give us just one or two more clicks of zoom please just so we can for those of you Yeah sure yeah just to make sure that people can read it because it depending on the connection sometimes it's it can be a little pixelated so that that that's pretty good that's a little bit better that's nice yeah much much better um yeah

so basically because of this it's actually a bit hard to pass and what we actually find with spreadsheets is people put their own sometimes human input their notes and that is actually sometimes useful context but sometimes there is no context um so really we actually were trying to find if there was something else out there we couldn't find it and we felt that the solution

was actually to build a tool where you can upload a spreadsheet you can chat with it explain that this is what's happening and then it can actually then turn it into more usable data that can be used for loads of different um pieces of of uh L&Ms and and agents and so forth um our plan is to sort of actually

open source this um because yeah we're sure potentially help a lot of other people um so at the moment we've kind of maybe something a bit easier we started off with a sheet like this which essentially has all financial data just shoved on one tab on one sheet now this is probably not the best you know it's got the P&L the balance sheet and the cash flow you look at the financial years these aren't actual dates

um so I'll pass over to Muhammad hopefully our demo works uh we're calling it Project Bloom um I quite like the name Blooming Sheets it's quite quintessentially British um but yeah I I'll pass over to Muhammad who can sort of show you what we currently have um and I think we have some questions for you lot as well yeah let's do it sweet sweet sweet okay um okay so we

we've repurposed one of our landing pages from a failed uh project before but um you know don't worry about that um so yeah so like like Kuan mentioned the idea is you kind of put in your your sheet here um and um initially what we were thinking about was having a voice agent and almost having a conversation with an agent and kind of um showing it exactly how your your

spreadsheet is structured you know the intricacies of the weird quirks that you might have included in there um I'm having a bit of trouble with the voice API so that's on hold for the minute that's tonight's job um but right now we do have a like a text version um so you put you press uh clean with text it

starts generating some questions uh based around the things that it can see directly in your spreadsheet um so if I just let that happen there we go um so yeah just ask some clarification questions um the idea is that it it knows that this spreadsheet is trying to be turned into some sort of database

table um it needs to be database ingestion ready so the questions are kind of based around that um so what I can do is just answer these questions um they're a bit arbitrary um yeah I don't think it really should matter um let's go crazy here and so these questions are generated then from an LLM correct exactly yeah

and is that just a call to a master agent that's doing we have a master agent uh using an MCP server uh that we actually found on Smithery uh funny enough um but it's yeah it's hooked into an Excel spreadsheet and looking around it and um then uh generating these questions um yeah maybe I can let Haroon answer just speak some ramble while I Yeah yeah

basically I'll say to Muhammad the the plan is definitely for this to be a voice agent because what actually happens is it takes a spreadsheet it does an analysis then it outputs what it thinks needs to be updated um I think I told Muhammad listen to just use something like Life Kit but he was quite adamant on on sticking with Mastro trying to get it work so um after if we

we get some time then we can maybe just look into how best to actually get the voice element um working but it'll be conversational and it will pick up maybe potential issues that are coming in the questions and it'll allow the user to input more information and that's kind of what we found is actually required to get this kind of information more usable

cuz as I was sort of explaining before sometimes there's like This is sometimes like hieroglyphics you know people put all different types of things on their own spreadsheets and you know an agent that you can converse with and actually get that information out is we feel is probably the best approach to actually

structure this information um yeah and I think it looks like M's nearly there with with all the questions yeah there we go um so I can hit continue um basically just uh give us an overview of our responses and then from here we've kind of add added two different options um I'm sure you guys I think you guys do know that the Superglue folks um we

spoke to them actually a while ago about this kind of problem and um and how they're trying to kind of tackle it themselves and I think obviously they're going after a lot more structured data and API data um but it was interesting that they produce kind of a specification um internally on how basically all the

data gets mapped into the required output format um so we're trying to do something similar where we have kind of a proprietary specification that allows um kind of an agent to actually go ahead and perform a transformation in a deterministic way on a spreadsheet so you can keep doing it over and over

again and almost turn your spreadsheet into an API um so that's kind of the where we think this this could go um but if I hit this fingers crossed this works this invokes another agent um that is more kind of designated to actually building out this specification but again it has access to the same MCP

tools um and hopefully this should work there we go okay um so yeah it's basically it creates this specification and I'm pify prettying it here um but you can see it's going to split this that spreadsheet into three different tables um each of them have their own um kind of quirks so I can hit apply transformation and now in my

uh transforms they have been uh created so it's kind of done some some cleansing here split them into three tables it's quite small but yeah it's a little bit kind of iffy um we're trying to kind of figure out the best ways to provide the most amount of context to make sure that it's kind of rock solid um but that is that's kind of what we have right now um

I can also show you the uh yolo mode so this mode is a bit different it's kind of like a oneshot you just hit you give the agent this information and the spreadsheet and it kind of just tries to figure out itself um so we can try this again this is a different agent as well um and we can see uh what it tries to do

and so and so in yolo in so yolo mode it just it doesn't ask questions it just goes is that is that right exactly exactly a and so so I guess let's I'm really curious on kind of like so the use case of this I have the have these spreadsheets with these unstructured data and I want to basically get these

spreadsheets in a format that I could maybe export to a CSV and import into a database or something right is that is that kind of the the vision of this so it's like a transformation tool for from spread from spreadsheet to database table yeah exactly um I think you know what petals does is the whole point of

pedals is it brings all company data from different data sources together now spreadsheets is a bit of a a challenge so as you're saying it could take that unstructured data and then you know pass it into the database or use it as a CSV but even just general use for LLMs like sometimes if people I'm sure maybe others have encountered it you just

upload it into GPT if the format's really bad it's just not going to work you know so even those kind of use cases in terms of allowing LLM and agents to actually pass spreadsheets and make them more usable is another thing we kind of see where this is quite useful for and this kind of stemmed this kind of

stemmed from like a personal frustration cuz we we when we actually on boarded our first client they gave us one of these spreadsheets and we had to spend an hour with them just basically them just describing us exactly how this spreadsheet works all the macros were everywhere it was just it was a complete mess um and so we thought you know this

actually could be something worth uh worth playing with when we were building Netlefi connect um and we were taking these enterprise deals on or trying to right in the discovery phase you we would always notice that legacy enterprises 20 plus years in the business they have like five six content management systems over the years and one of those C CMS's

is a spreadsheet or a Google sheet that somehow has been maintained for 20 years where they add content to it and it gets synced into some like process right which is ridiculous by the way um so I do see the value here for sure yeah 100% i sort of have a background as a data consultant and I've worked at like loads of different companies analytic with Google and and even still

like yeah people just end up using spreadsheets as CMSs CRM all sorts you know yeah yeah i mean it's it's you know the question is I think people ask themselves why can't we just do this in a spreadsheet and then they quickly realize you know and spreadsheets are great for some things right but oftentimes you start with a spreadsheet and it makes sense but then years into

it it becomes a monster and they want to move off the spreadsheet it's not like they don't but they can't because this crippling amount of work to relocate right and do all this stuff so another plus one for this exactly yeah well this has been really cool um I appreciate you both coming on i'm excited to see you tomorrow at the you know at the awards

ceremony and I guess any any other like closing comments or or thoughts before we we bring on the next guest no you know just uh I think man's like a super master fan so just uh thanks for thanks for having us on and um yeah we appreciate the time yeah we'll definitely have to have you come back on when as you get even further and and show some progress and do an update and

I think people people like seeing this this is what we want to do is we want to show that all the different types of use cases you can build with AI agents and I think seeing people in the hackathon building real things is is just a way to hopefully inspire a lot more people to see that they can build interesting things as well so thanks for doing this thanks for coming on yeah I know you all

had questions about voice and stuff let's just talk in Discord i'll see you guys there yeah see you guys all right dude we got 89 of us in this chat right now yeah we you know we are uh Turns out if you just live stream forever the numbers just keep going up that's the hypothesis yeah 24-hour live stream coming up next uh but yeah we did I did want to bring one more uh guest on so today we want to

bring on a guest from 3:25 now 3:25 one second now we're at 3:25 okay welcome the 20 people that just joined yeah so for those of you that just joined I'll do a quick recap of what we've done so far and what we're going to do to kind of close this thing out we're getting close to the end but we are not there

yet so I started this thing off by showing how to basically architect a Nex.js and master project some different options we talked through the trade-offs all these things are whether you're watching on YouTube X LinkedIn you can always go to our YouTube channel and see these things after the fact so if you want to see what you missed you can go back we had an AI security corner with Ally that she was a guest that

joined us to talk a little bit about security and uh how you should think be thinking about security when you're building AI agents we talked to two different groups building projects for the master.build hackathon so if you have not heard about that go to master.build you can learn a little bit more we have the awards ceremony

tomorrow it'll be live streamed starting at 10:00 a.m pacific i wasn't expecting Obby and Tony who's somewhere in that room i wasn't expecting them to join today because they're in Japan and it's like 6:00 a.m you know so joining early you know no sleep for uh no sleep for you and but last but not least I did want to bring on someone from the Mastra team so we're gonna bring on Shreda and see what we're

gonna talk about for the next uh 15 or 20 minutes or so hey guys what's up Shria not much just chilling i can't believe you're on Obby i'm so jealous that you're in Japan it's like my bucket list country to go to it's pretty nice here there's no trash cans on the road so there's no like you have to actually find a trash can which

makes you not want to consume anything outside it's kind of a trick i like it interesting uh so Shita maybe just a quick summary of your background and what you do at Mafra and then we'll talk about what we want to do today as we kind of close up this live stream yeah sure um I would say that I am a writer um I studied writing in CS in college

and somehow that ended up being like the perfect combination of things to study for my career i was previously at Mercury um we have a magazine called Meridian and I was the deputy editor and just helped get that off the ground i wrote the profile on Dwarfish Patel which I'm sure a lot of like AI peeps uh

know about him and love him um and now I'm at MRA i just kind of noticed that this is the direction that tech is going in like you either adapt with the AI wave or you don't and you like perish I guess i don't know it's like then maybe it's a little too hardcore to say um and I realized like oh like you know I really did not actually enjoy building uh web apps in college uh but like with

this new I guess wild west or like new like programming renaissance it's so much easier and so much more fun to actually do it um and so like I don't know I feel like I am actually going I I want to become an engineer I guess uh or like a content engineer which is I think the title that we arrived at um role at MRA yeah it's it's like part writing

part engineering role and how do you how can we use the tool the new tools of the day to think about what what should it what are what's the next stage of writing look like yeah exactly i I mean I've been reading um uh this blog by this author that Sam Sam and I actually met through reading his work called Veniteshra and he um you know I had no

idea what how he would feel about AI but uh he actually has been just like writing entire blog posts with it and is saying that um like don't writers need to maybe get less attached to the craft like the object level craft of writing and focus more on like the ideas and concepts themselves um so I I find that

you know like that's exactly what's happening here with Monster is like you can be so much more like fast and produce so much more content but uh you don't want it to be slop right like you just spend a lot more time making sure that the actual quality is good and I would say um a second observation that goes with that is like the craft has kind of shifted into prompts like prompt

engineering like that is the new craft actually it's like the meta craft so I guess we're just moving higher up the abstraction chain yeah yeah yeah i guess my this is maybe my hot take but I think uh prompt engineering slowly going to fade into just being communication right because you early models you had to talk

to it in a certain way to like engineer results but I think as models get better it's going to just come down to clarity of ideas and clarity of communication so as long as you can communicate very clearly I I think the structure of it is going to be less and less important and the quality of the communication and the

ideas itself are going to become more important so but yeah it's it's kind of an interesting you know interesting wave we're on for sure yeah exactly you said um that you didn't like to do like webdev and then now this is like super fun it's actually the same thing that happened to like Shane and I when we

were doing webdev and the previous in or the previous regime was like serverbased apps like with PHP and stuff the minute you can get to this iteration loop where you get like instant dopamine from hitting save then you becomes fun again right so that's what a lot of people are feeling now exactly honestly just to

reiterate that point one of the things I I loved about Gatsby and the thing that like the reason I jo ended up joining Gatsby and meeting Abby and Tony is when I was playing around with it the first time and this was kind of early into Gats you know Gatsby's existence but just the live dev refresh the idea where I could click save in an editor and not

having to refresh the browser that was pretty miraculous i'd spent you know years building Drupal websites and PHP based websites and you know it's not like a refresh is that hard but if you re if you're building a site and you refresh it 400 times that day it's kind of nice when you don't have to click that button you know going from 400 refreshes to two the whole day you know

it's so it's that iterative that iterative loop and so yeah I think I'm seeing a lot of that this cycle as you can get to you can get to that feedback loop even faster because AI helps you get through that prototype stage or helps you kind of like get through the the things that you would normally just have to spend a lot of time hands- on keyboard to get to yeah I actually

wanted to see if I could uh share something that I literally just built prior to this call um so obviously as a writer here at Monster First of all can you see my screen yes but you're gonna have to give us like one or two clicks of zoom okay great is that better yes and if you want you can also hide that

left sidebar if you click on the right in the middle right on the the border oh like right here yeah just click it okay great a little more room thank you um yeah so I actually literally just built this tool uh where I'm going to try to prompt um my agent and get it to automatically draft uh a tweet uh for Sam's Twitter um so I'm going to just

pull something from our uh recent change log which is that we uh support agent to agent communication um so I'm just gonna say something like draft a tweet in Sam's voice about MRA's uh latest I guess at MRA's um uh support of ADA communication okay let's see if this works okay so it actually um it actually I

guess looked at this MCP tool that I gave it to like try to get uh some knowledge about Maestra let me see if I can get it to uh do it actually in typefully it was working earlier okay so it's using this tool it's using this Zapier connection it says it's been uh drafted successfully um let's see i'm going to actually like refresh this sometimes it

takes a few minutes for it to show if we if we go back can we look at if you expand the tool does it give us any information about what happened it just says it's saw it actually in the in the typefully list oh was it there success is false never mind i'm sorry i thought I saw it in the in the type playlist like if you go to the bottom left Oh go the next one oh yeah here it is

it's literally here yeah you got you just have too many you're drafting too many tweets for Sam yeah exactly no I was testing this so that's why it's over here um Sam Sam has a lot of tweets to review and edit exactly um this is really cool i just created like a web hook uh with Tyfully and I guess I could

show I wonder if I can show some of my actual Monster code yeah let's do it i mean I think the nice thing about this is you are just connecting tools right you didn't have to write this integration to typefully you just knew that okay that's what we use that's what I use to draft my tweets that's what Sam uses you obviously help us with with

some of that and yeah we can just connect it to an agent that's that's really cool i'm having this dumb like uh system settings share issue sadly i can I don't know maybe I can just uh Yeah this is not going to work unless I like restart Chrome sadly but anyway I was able to do it maybe we can publish this as an example or something yeah well let's yeah we'll we will uh you

know follow us and we will publish this and put it add it to our examples in the repo because it is kind of a cool use case of really simple way to connect tools to accomplish a task right like that that's I would think that you know no matter what whether you're building tools you know software tools for others to use or you're just using things internally like there's really good use

cases for small kind of verticalized or specialized AI agents or you know AI workflows and I think this is a really cool use case of that so all right uh so Shria how can people follow you I guess I'm literally uh free shr ITA on Twitter i have one of the Mastra affiliate icons near my name now so that's cool all right let me drop

that in so people can follow along if you if you want to see what Trita is doing yeah I'm going to actually write an article soon um about what a content engineer is and it's going to be published i won't say where but it's going to be published somewhere so I'll tweet that out and I I guess my last point is like I think this is probably going to be a new role that

a lot of companies adopt it's like the evolution of the content strategist or content writer yeah that is uh I I would agree i think that you kind of can especially when communication is what drives a lot of these models right and a lot of these models can write code there's there's this kind of blend of good communicators which I would say writers predominantly

you have to be a good communicator can now also become and start to become good engineers too right you know there there's of course you know nuance to all this but I do see this as like a pretty major trend all right thanks for coming on thanks so much for having me great work by the way thanks yeah yeah see

you dude this was a sick episode dude 362 viewers crazy i mean our longest stream right this is our longest stream i mean I feel like we got to wait we got to be You and I got to hold it down for eight more minutes so we can say we did the twoour stream dude we will definitely because I have I I can tell you some things yeah well yeah I definitely want that so we'll introduce another dude he's still

live yeah you want to come on yeah we need We need We need to We have to kill eight more minutes so we can say Agent Ashman ashman dude what's up how's it going i was just telling Abby I haven't brushed my teeth yet so I don't know if I should be on camera it's 6:30 in Japan this is what happens luckily I cannot smell your breath from here so

you're good you're good dude yeah okay sorry dude um yeah well uh yeah tell us Ashwin maybe you just started at uh at MRA what what is your first impression and what what are some of the things that you're going to be helping us out with oh man uh well first impression is I'm so excited to be here this is um such a such an energetic group of people

and I think a time uh you know just in tech in general and uh you know it's funny i I didn't know I was coming here to Japan on Monday night and then Wednesday afternoon I was getting on a plane we make decisions quickly yeah yeah and uh you know first thing we get in and I'm like "Hey guys I I have a meeting I want to take you guys to like this was at like 10:10 last night and we are

hopefully going to be doing some cool things with a pretty big multinational on Monday." And so things are moving fast uh and uh you know I Yeah just exciting so exciting you know I um I all the books are over here like Sam had actually sent me a copy of the PDF before it went to went to publication and so I'm Yep there we go i had shared

that PDF on a lot a lot of different groups this is the book of Mastra you know you got Everyone's got to have Hey if you want if you're watching this if if you want the book of Mastra you can get it it You can get the digital copy for free just go to master.ai/book if you want the physical copy here's how you can do it it's very

easy we have a hackathon going on right now submit something you'll get a free copy of the book otherwise if you go to meetups that we're going to be at in San Francisco Ashwin will probably be there and he'll probably have books so there's a lot of ways to get this book you can also buy it on Amazon if you if you can't wait you just you want to get the copy you can buy it there too i I

literally brought 60 pounds of books to to Japan so if you happen to see me wandering around Tokyo I can I I'll give you a couple copies of the book if you see the three tall dudes taller than most people we got books okay yeah luckily Luckily you can't tell on this the stream i would I would not fit in with you three probably but you would fit through the door actually is perfect

height for most people here tony hit his head like walking through the door he hit his head like because he didn't look Yeah i mean yeah it's just different different strokes different different culture yeah different different heights apparently yeah dude for sure yeah yeah see you Ashwin have fun i keep forgetting it's 6:30 in the morning here you know and I am just

wired awake well I suppose your time zone change is is pretty pretty intense this might sound more ignorant but like there are some cool things that I've just discovered like there's like vending machines outside that serve coffee and Coca-Cola and all these different drinks so like on your way to work you know you can just walk out tap your card get your coffee and go it's so

cool and it's so clean here i love it like it's just there's no trash anywhere i love that and people don't jwalk which I'm such a jaywalker in California but everyone is like just waiting for the light and I'm like how come no one's just crossing right now like there's no cars coming but like there'll be like a

herd of people if you're in New York or California or anywhere you you know people just jet across the street who cares you know um so there's Yeah cool stuff yeah well you'll have to find some some of our you know Japanese super fans and we can get get them on a live stream or something as well i know I know we have just a ton of community support

there and that's why three of you know three of the mastros are there right to try to try to meet people try to you know I don't think we really predicted that we would uh like we didn't predict that Japan was going to be this big thing right there's nothing that we did specifically to try to reach out to Japan I think there was a couple things

that probably helped and then uh we we've of course kind of saw that early and just now we're trying to just really really do what we can to to help we translate our docs to uh Japanese if you go to our docs page you can find the the language switcher and you can read our docs in uh in Japanese so trying to do

things to help bolster our community there because it just is so strong yeah we started following all the Japanese uh tweets that we get too and try to respond to those there's so many every day so many yeah like we we have this so we have this tool to try to keep track of who's tweeting about MRA or us questions so we can try to respond and

it was only filtering for people that use like MRA AI tagged us specifically in English language and then we realized there's actually dozens a day it seems like sometimes more than that of just uh Japanese tweets and thankfully uh on X you can click the translate post because otherwise I would uh I'd be screwed yeah

it's been good feedback too it's not like oh like people just saying you guys are amazing it's actual good feedback there yeah I mean yeah we've been pointing out that a lot of things are broken or a lot of things aren't working you know we we've you've been doing this what six months now or so a little little more than that seven months that

we've been working on MRA and things are moving really fast and so for those of you that are tuning in if you have any thoughts or questions you are probably watching this on LinkedIn or YouTube or X just drop us a comment and we can answer any comments as we start to close out here we've talked a little bit about MRA with Nex.js we talked a little bit

about AI security we had some people come on from that are participating in the master.build hackathon and we talked to I guess two people from the MRA team Shida and Ashwin yeah it's a jam-packed stream today yeah it's it's honestly just busy there's so and tomorrow's going to be another may not be quite as long tomorrow but just a lot of people wanting to to get on

tomorrow we have uh two different friends from YC companies coming on one to promote another hackathon and one to show off kind of a new project that was built so we got we got Eric coming on oh Eric's coming oo that's going to be so sick yeah Eric Eric's coming on we're gonna we're going to talk a little bit about Pig we're gonna I'm gonna talk I want to talk to him a little bit about

banana because that still holds a special place in my heart if you don't know what those means you'll have to tune in tomorrow and find out exactly exactly dang that's gonna be that's gonna be a dope stream I'm excited for that and yeah also if you are building really cool things with MRA you should be in the hackathon even if you're not if you're building something cool you want to talk about it you want to share it with the world let let us know you

can find us you can find Obby on X right there x.com you can also find me x.com smthomas3 shameless plug if you're one of the 374 watching this and you have not starred the Maestro repo on GitHub what are you doing with your life please help us out don't be a DJ yeah we we would appreciate any stars any you know

any help you can provide and honestly if you are using Maestra and you're running into problems please go to our Discord please let us know we do our best to try to respond to every I mean we have we introduced we tried to introduce like an SLA for ourselves to always respond within a certain number of hours to every Discord message we're not always perfect but we do try to respond we do

try to make sure that if you are running into problems we do our best to help you and honestly this is that's our best source of feedback of things that are broken things that need to be improved and uh yeah honestly just we learn so much just from the things that you share with us in our discord so please join that the hackathon has been very

illuminating in a lot of ways because we've started this in October and I did like on my way to Japan I shared with the whole team like this like trip down memory lane and like our first like thing for doing this was like October 1st and we've grown a lot and like but we have so much more to go um and so if

you're following us like yeah we're going to fix all the bugs eventually you know we're gonna figure things out more so just pump the stock while you can yeah and and tell us what we need to fix so we know we know how to prioritize it that's what the hackathon really did it showed that I mean as as like as us as like creators right our whole team as

like the people building this thing you never really get to use your product as much as the end user right they're building these cool things and like all the things we see are great but then the hackathon also illuminates where we suck and that's where we have to get better and that was dope so thank you everyone i mean it really consolidates a large

group of people starting from the beginning at the same time and so you the whole like getting started experience it puts it under a microscope and then you you see things like our story of how frontends connect to MRA needs to be tighter needs to be better yeah it's just not good like there you know we hold ourselves to pretty high standards so we you know I know a lot of people if you're watching this like you

probably are like "Oh MRA is cool like the docs are great." We internally we're like "Our docs suck yourself you know our our our story around our examples around how to connect your front end to MRA it sucks so we are you know we're pretty harsh critics on oursel and so but we only can get better if you tell us so please yeah join us in Discord

chat with us let us know come on live streams if you have cool stuff to to share and if you know anyone else that's doing something cool in the AI agent space please make some introductions we're trying to uh showcase all the cool things that people are doing and this is one of one of the ways we do that

we did it two hours yeah we we made it dude two hours first two hour live stream this is the most I think that's ever joined right this is the most we've ever had 374 now 38 for being here yeah so we 383 i feel like I feel like if we stick around for three more minutes we're gonna hit 400 and we can celebrate

should we should we yeah yeah we got we got nothing else like we went through you like the new like icons on the bottom it looks Yeah man i go I go on a trip and I come back and like we have all this technology now yeah we ask ourselves every day how can we level this up just a little bit i mean we you know I hope it looks

somewhat polished but we don't know what the hell we're doing so we're just figuring it out you're hopefully you're all uh figuring it out with us and uh this is uh Jeremy Alston i want to share it with the world please do yep please do this is kind of funny i'm not even gonna say it out loud but I just laughed when I read it so that's hilarious

a lot of people feel the same way yes um I would say I would say that is so yeah let's just go ahead and wrap up this has been a great stream we got two hours we we did a lot if you want if you just are joining and you want to see what we covered go to go to our YouTube channel do we have that somewhere where I can

share it we don't i'm going to find it but you can find us on YouTube and all of our live streams are there we'll probably try to like cut some of the SE segments up and release them on their own because I think individually there's some really uh really interesting things so the master YouTube I'm going to

display this on the screen and then we're going to get the heck out of here you can go there uh right there find us on YouTube MRA-ai you can see all of our past live streams we do this every day Monday through Friday at least we have been for the last week or so and we plan on continuing it normally it's an hour

today it's the we got the bonus hour in so feel like uh feel like I need to stand up and maybe stretch my legs same car dude