Daniel from aiagent.sh, AI news, Meet the Mastras, YC X25, building a DJ Agent
Today Abhi and Marvin chat with Daniel from aiagent.sh, they talk through the latest AI news, discuss YC X25 companies, meet with some Mastras, and build a DJ Agent.
Guests in this episode
Episode Transcript
what up you guys excited i don't know if you want some light on oh yeah sure i mean we already started so what up welcome to AI Agents Hour we're just getting set up here um yeah hey guys how's it going i'm Abby i'm Marvin again again uh so if you came yesterday uh we are still in Strawburg France hometown i actually
leave today from here to Brussels to meet some homies and the world tour continues um from there I'm going to a wedding in Greece of my friend Tony from Austria so shout out to Tony for getting married and uh then I'll finally be home i'm going to be missing the AI engineer World's Fair uh which is sad I believe
but uh You should do that once a year or something like that uh so we'll talk about the world's fair hello SF are you from SF i don't know um yeah so today's show we have some AI news like usual we are going to be uh Salo Leon he's a friend person that I know oh you know him what up Florian that's cool um yeah
coffee we're also going to have some friends from the community um Daniel he will be here and I just want to make sure I have all this information we're going to have one of our community member friends Daniel Trevino on after the news then we are going to bring on some mashras that um have learned some stuff that they want
to just kind of share and then yesterday uh while Marvin and I were eating tart duom and drinking some wine we're thinking like could we make music with AI and that's what we're going to try to figure out so we have a DJ Mastra the DJ Mastra today uh so it'll be a fun show once again 31 people here really appreciate everyone here so uh let's get
into the news also I'm going to send the invites to everybody who is going to join us so I'll do that here oh SF is actually ah I did not realize until Bonjour SF bonjour um cool just setting sending links to people yep but uh Yep i'm drinking the coffee today yeah dude coffee hits different when you're doing live streams too okay I think we
all set up and we have to we've told the PE powers that be when to come and all that uh we'll we'll kind of introduce them um but before we begin uh Marvin where are you from in France so I'm living in the frontier of Germany and I come from this small city that is mentioned in the chat which is so yeah uh I'm at the frontier of Germany right now i have mountains close to my house like you can see them by the
window and it's Germany actually which is a funny little story but yeah both nice we have another guy from France Amin hello from Paris is trash though honestly it's super nice man it's super nice I literally started learning French when I came to visit Marvin And I was like living in Paris for a little bit i mean and uh
when I came back to Europe this trip it was crazy how your memory we're going to talk about that today um it's crazy how your memories can like re-trigger and like all the words that I like the words that I was looking at I started like coming back like "Oh I know what that word means i know what this word means." So anyway let's get into the show um
we're going to start with some AI news let me uh get it popping for us um so here we go is there any BAI today or no BAI no IBMs no SSGs whatever um so first things first we um we're going to play this video super interesting like this is not new news at this point at all but I wanted to share with everyone like uh this kind of open
AAI acquiring Johnny IV startup and like we can all ponder what it means um and that would be really cool actually if the chat if you want to think if you also want to draw some conclusions we're going to as well um but I figure also need to make sure I shared the sound one second but we'll watch this together and then we'll like draw some
conclusions um where's my share button oh it's right here it's in the other tab so we don't know where I am here I am let me stop sharing need to share again share audio okay great we're good all right let's have some fun and watch this io is merging with open AI formed with the mission you got ads okay hold on and so IO is merging with open AI
formed with the mission of figuring out how to create a family of devices that would let people use AI to create all sorts of wonderful things in a big move OpenAI is acquiring a year-led hardware startup called IO Products for a staggering $6.4 billion as per reports but this isn't just any startup it was founded by Johnny IV the legendary designer behind Apple's most iconic products this is OpenAI's biggest
acquisition yet and it signals a bold new direction building AI first hardware not just software during his 27-year career at Apple Iv was the mastermind behind the iMac iPod iPhone MacBook Air Apple Watch and AirPods known for his minimalist aesthetic and obsession with detail I was kned in 2012 for his
contributions to design and innovation after leaving Apple in 2019 he founded a creative agency called Love From bringing together architects artists engineers and designers a year ago IV and former Apple colleagues quietly launched IO a hardware startup focused on devices for the AI age now OpenAI is
absorbing that entire team openai CEO Sam Alman and Johnny have have reportedly been collaborating for 2 years their partnership evolved from tentative ideas and explorations into tangible designs the details of those designs remain under wraps but the stakes couldn't be higher so why is OpenAI betting big on hardware because
the next frontier of AI is physical devices that seamlessly integrate generative AI into everyday life outside the confines of smartphones or laptops two years ago Johnny and I started talking about what the future of AI and new kinds of computers was going to look like i was running OpenAI johnny's running a design firm called Love from that had established itself as really the I think densest collection of talent
that I've ever heard of in one place and probably has ever existed in the world and it became very quickly apparent to both of us that we needed a third company open AAI doesn't want to be just another app inside Apple or Google's ecosystem it wants to own the hardware platform that defines how we interact with AI much like how Apple defined the
smartphone era analysts say this is the same ambition that Meta has with its Quest VR headsets and Ray-B band smart glasses startups like Humane AI and Rabbit have tried and stumbled humane's AI pin despite being founded by exapple talent face criticism for poor battery life overheating and lack of real utility open AI have hired hardware and
robotics experts including CKNowski formerly of Meta's AR division and invested in robotics startup Physical Intelligence the message is clear open AAI wants to bring AI into the real world not just your screen i won't become an OpenAI employee but his firm Love from will now lead the design for all of OpenAI's products both hardware and software their teams will work closely out of San Francisco where
Iive is also designing a new headquarters for the venture meanwhile Apple which has been slow to roll out its own AI features saw its stock fall more than 2% after the news open AI backed by giants like Microsoft and SoftBank now valued at $300 billion for more news and updates keep watching Live Mint
ringing the bell okay so 66 of you just joined 67 now well welcome to the show oh no back to 66 someone just bailed you i'm just kidding it was me dude so OpenAI getting into hardware johnny IV not joining the company but essentially giving the talent um that's actually I mean good good commentary let's put you up first some folks say it's a dystopia i
don't know why yeah well Chad please put your thoughts too on like what do you think like what do you like let's all ponder like what does this mean or what could this mean couple things just from my top of my head like 6.4 billion dollars to hire talent that is probably really good um that's a lot of money dude it's a lot of money hardware
hardware engineers robotics engineers are expensive talents right so you got to think like I don't know how big their team was maybe we can look see if we can find some information on you mind like searching if we know how many people are on the IO team but uh that's like let's just say you know a million dollars per head i
don't know how many people are on there but that's a lot of money as it is then you have like you know company valuation etc yeah but Johnny's not joining that's just interesting so it's like it's almost like he took the money um this is okay um this is more pessimistic takes right i have some optimistic takes too pessimistic takes are a lot of things in
Silicon Valley are incestuous um and this seems like another maybe this is another one 55 member team right okay so let's say your total comp 55 people is a million dollars each right let's just say so at 55 million in salary and stuff but then you have like valuation of their company and so 6.4 billion is the value like I mean that's the deal price you know and
you can imagine I don't know how much revenue these guys were making before i think they were like a design firm so I'm sure they were making money though because like you got like you got the Apple name on it um and probably a bunch of like alumni from Apple in there so you could see how the price will go up
very quickly if you want to hire talent you're not paying a recruiter for this right you're just getting a team of people that have already worked together i'm assuming it's kind of like MRA like already worked together have been doing it for a while already know what they're doing they're they're probably already working on something right so this is this is
interesting and incestuous of course right um what do you think dude yeah I mean this is more the OpenAI adding hardware to their to their fleet that is a bit not annoying but a bit weird to me they're good with dealing with data huge volume are good at building I mean LMS building the models providing them to us and now shifting to
hardware feels like I mean they have money they can buy pretty much everything but it feels weird I would prefer them to focus on what they know that what they know how how to do things instead of focusing on this kind of new new stuff it's like I wonder if it's it was just an opportunity like meeting this guy and saying hey let's go we can we want to work with you we do want to do other stuff and yeah but I'm
not sure I don't see unf doing the same move for example yeah maybe maybe not or maybe you have to right or maybe open AAI is trying to go after you know Meta's market share right so they want to have their own hardware devices obviously the quest is okay um so also like another I mean dude you're killing it with the commentary bro like this 100% msure could it be a strategy to
lock talents and avoid with them yes that is exactly what you're supposed to do when you're doing cutthroat business right especially if these guys are talented you're going to pay them a lot of money to work at OpenAI it's not like they just got the money and then they get to leave right they all have to work there for a couple years so I mean a lot
of people got rich and they're not going to go anywhere for quite a while i'm pretty sure they'll probably have golden handcuffs uh for any of those joining us if you don't know what golden handcuffs are I don't know how many of you have been part of startup acquisitions i've been part of two um every time you get golden handcuffs if you are someone that
is going to the next company and some companies will do either six months a year and but as your compensation goes up from the acquisition the golden handcuffs go up so the time that you have to spend so my first company was called workpop like my golden handcuffs cuz I was like 20some young I was like a
young guy on the team my golden handcuffs were zero they paid me and they were like "Hey if you want to stay stay." And I was like "Nah I'm going to Gatsby." Uh and then when Gatsby got acquired by Netifi there was some compensation involved there but it was a six-month handcuff um and then there was
like a sixmon other six-month period if you hit a bonus if you hit a year you know it's all to incentivize you to keep talent in one spot and not go off because if you're good enough to start a company you're going to do it again right 30 that's how entrepreneurship works and if you're good enough to do it again you
will so that's why they have to keep you there so yeah that's just all the financials on that let's talk about the good stuff like positive piece of this if AI is built into hardware directly innovation man super innovative but I mean Tesla's already been doing this so next level stuff this is like some optim optim optimism optimist or whatever um
but you're already way ahead of us like for for example you mentioned Venmo yesterday never been inside this car and I'm so I'm so scared of just you know putting my feet in a car that drives by itself like wow do they have Whimo in Paris or anything just in America i don't know maybe some people from in
Paris are in Paris in the chat and can tell us but I don't think we have i don't think it's legal yet in France yeah so it's not like it's you know like the point here is it's not like no one else is working on this right Facebook's working on or Meta is working on this you know and even some some friends from
my YC batch are working on like the the meta the meta smart glasses and stuff like that um so yeah it's just an interesting what a time to be alive is what people would say we'll see a lot of new things coming in the next few years be crazy so on to our next topic uh let me share um I'm going to share this so we already shared this before um I wasn't there for
that stream but um uh I want to talk about muscle memory uh so Eric did come on live stream i think you guys can check out that episode i believe it was last week or something and um yeah and so um I just want to share this with everyone muscle memory removing LLM calls from agents so funny story behind
this is you know Eric is from a company or the company is called Pig he was previously the founder of a company called banana.dev they were doing GPUs uh like a GPU kind of platform and then during YC Eric was doing PIG which was like Windows automation um which is now it's like RPA which is remote remote something or I don't know some process autom process automation in
Windows and as you if many people don't know like Windows apps like run a lot of important things like hospitals and whatever right these and they're kind of older systems too like Windows XP are still running or Vista and like that so what he would do is you would spin up these virtual environments right and then you can run automation tasks by
spinning up the virtual a and then this he had an agentic component which he's like taking snapshots of or screenshots of the page and then you know you're going you're like automating processes so if you like from the United States or something there's this like a program called Epic you could start automating
some old software that's maybe not on um like a SAS right and uh since then we were talking about this at like a sushi dinner and like a lot of this work of doing automation is a lot of it's unknown in the beginning right like when you need an agent to discover a page like for example I just drop it on a random
application much like a human right we would then start poking and clicking buttons and if we get confirm dialogues we'll read them but maybe we we don't want to confirm anything we hit cancel then we go and see what's in our profile tab like we explore the site right but now when you go on Instagram or whatever
app you use you're not thinking about where the profile button is like you just go there Right and with automation right like if you've done this automation process before in the sense that you figured it out you went like as an agent agent you took this you did all the the research and discovery to figure it out why do I have to pay for tokens
again like I've already created this automation like I should not have to do LLM calls again for this reasoning of discovery this is where muscle memory comes in right muscle memory is is essentially just like that in humans right after you can figure out the patterns for tool calls that you may need to make then you don't need to do it anymore right you turn this into a workflow or an executable that costs you
nothing because you know exactly the steps you have to do um memory in general is just such a fascinating concept but I will post this here for y'all and um I think what we're going to start doing um in the future is we'll have some like show notes after the show so you can see all these links that we post cuz you know may be hard to get
these so okay we got to move fast because I believe our guest is here so I'm going to go to the next the next topic and I believe this is the last couple topics here um next all right next up we have a tweet from our buddy Showya he's from a company called Jinba in Japan um and he's and Jima is like a MCP tool uh so
MCP workflow tool um but I love these market maps so I think it's just really interesting for people to see these market maps because they've been expanding and changing every time I look at them so let's go through some of these because before we didn't even have MCP marketplaces like almost six months
ago and now we do and there's definitely people missing from these things like always people always like missing some and then the the people who are like us like when they do the agentic framework stuff we're just like hey you got to add masha they'll never do it and it's probably the same thing for a lot of these um but let's start with the
marketplaces so Clavis AI is in the current YC batch so they made this this board which is funny because like MCP runs not here so that's whack but there's like more here composio's here their marketplace uh Smithery the homies uh what up Henry um clients duh Claude um Libra chat I've heard a lot of good things but not really libra chat I've
heard some stuff about it um clients cursor makes sense but then you know these are the idees client is like a big one copilot is whack but people still use it uh Replet's tied never heard of zed though it's something from the community like from the view community I would say oh really like it's an editor
i don't It's from the view community i would say so it looks like this i've tried it once somebody in the chat might know about this michelle if you're here I know you're with things oh you know the audience dude some of the There a bunch of French people in the audience huh french guys yeah there's just another yet another
server-wise i feel it's weird to put servers on an MCP market map because I guess it just said states that they have servers technically anybody could have an MCP server so okay frame hey dude we made it we're actually on the framework thing i didn't even look at this we actually made it dang close to Have you seen the name wrong uh it's okay anthropic has a framework i don't
understand how like how does that work these are okay i don't feel good about making it anymore because it's irrelevant actually but uh Okay and then builders um we should be there yeah we should be gumloop solid make.com n Zapier oh is is it Zapier or Zapier we say Zapier zapier stack stack like Stack One or Stack AI so these definitely Oh
yeah jinflow so they're in their own thing and then hosting platforms sure it's just cool to see these market maps but in terms of like what like what this is it's really the servers that are the the big the big deal here you know so thanks for uh congratulating us on a uh not a really great market map but uh we'll take it nonetheless that's cool to get the same I mean the same stare as
enthropic can open yeah I just I don't know i'm a pessimist today good um okay then now I'm going to do a speedrun and then Daniel is going to come join us and then now everyone's going to be in for a treat uh because it's going to be a really cool demo um also I hope I'm not putting him on the spot by saying it's a
really cool demo but it is a cool demo so um I have to go through this thing we're going to be doing this a new new section that I just created 5 seconds ago which is a YC um we're going to do like a YC uh rapid fire so I'm going to just tell you all about YC companies that I'm looking at in our space and that you should probably look at them too and if
you're an investor I guess um this is not financial advice by the way is that what you're supposed to say this is not financial advice um but hey if you're an investor maybe you should take a look at these or um also YC companies are going to be fundraising soon so if you truly are an investor maybe I don't know i
didn't say anything um Bloom this company is dope because they are the mobile app that builds mobile apps which is honestly so needed it's like lovable for mobile apps so Bloom.ai AI or bloom.diy that's even cooler uh thing i've never built like I've built an iOS app before but I am not an iOS developer and just like people who are not web developers or
becoming web developers tools like this become really cool i remember like how much like how much time have we spent in our careers like fighting against oh like we'll use React Native and then it's not good and then you have to find an iOS developer and then it's like Yeah yeah it happens way too much to be honest i've been through this fire
multiple time and it's Yeah is this one generating like real native apps like using Swift UI or that's the that's the goal but uh how they do that they do they do use like sandboxes okay um actually they actually came over to the apartment one one day and we did like a kind of like architecture session so
they they are or were using MRA like I probably think they're not using MRA given the time they came over like we weren't probably as good as we are today um but they have this infrastructure that they then use code sandbox which a lot of AI companies are using code sandbox for sandboxing which is probably why together AI bought them so anyway I
don't want to spend too much time because we do have a time crunch i'll just go to the next one vibe lovable for internal apps this is a big um this is a big market here for companies who the problem with internal apps right is you don't really put effort into them and you don't maintain them properly because no one gives a right it's internal but if you can vibe code it and then continuously add
features with like less um you know less effort then this is cool it looks it's just like lovable I saw a demo at this at um where was it uh the future future of DevTools event i should know that because we hosted that event um next Casco agentic red teaming for AI agents if you want to add guard rails and you
know essentially protect your AI agents very cool these these guys from Casco are really nice they we I think we had beers with them so I don't know if you can have beer with someone you might as well invest in them all right last one keep going sorry um this one clarm zero hallucination AI deep research agents
this is becoming interesting too um so yeah if you're interested take a look at them then this is where things get more interesting ai healthcare agent for patients this is really cool varrage is a friend of a friend so I guess by reflexive property he is my friend now too um and they are trying to have an
ondemand healthcare assistant insurance companies will love this when you start getting sick people who need to do a routine to get paid by the government this is a US thing so maybe not applies for EU but uh very cool and we'll probably have these guys on later uh in another live stream sorry two more for rapid fire and then we're off to the races again uh Claude deep
research for people um so you want to go research people pretty sick dude um I could this kind of stalkery too yeah honestly little stalker oh my boy but hey if you ever wanted to deep research me don't do it but hey you can I guess um and then that's it for YC Rapid Fire we are going to now bring along our friend from the community one second let
me undo this and let's get Daniel in here so Daniel hey Daniel welcome to the Hello hello how's it going man so good thanks uh good to meet you and uh thanks for the invite so where you where you calling from yes i am based in Stockholm in Sweden uh but I'm originally from Mexico so um it's a bit of a mix here but I've been I've been in Sweden for more than 10 years now so starting to lose my Mexican where
in Mexico you know uh it's called Cornovaka it's like one hour from city okay on the car depends on the amigo gracias gracias yeah but uh thanks for the invite and I to be here so a couple questions before I put you on the spot um how did you discover Mashra and um I guess can you tell the kind of the audience your
opinions of it yes so um I'm a former Gatsby enthusiast and it just uh popped on my you know GitHub uh feed like oh some former Gats uh Gatsby engineers are building a m an AI framework and I've been exploring already by then with other kind of frameworks and trying to sort of build on top of them and the
developer experience wasn't really like there for me like every time I I like to be on on the edge of things like latest on my dependencies for example so every day I try to like bump everything test everything works and keep it on the latest so every time I did that on the other framework something broke and I
was like so like annoyed at some point i'm like okay you know what I need to give it a try to other alternatives and master hopped in and uh it was just so pleasant to the go through the you know documentation extensive documentation I still haven't uh got the hand into all of the corners so much to dig in also Daniel's being very like humble like he definitely used
MRA and he's contributed right he's contributed he's been okay with all of our bugs um and like kind of like worked like helped us and rooted rooted for us so thank you for that open source projects do not exist without users like you Daniel because um like you said like you're you're you're the type of guy who installs latest or alphas like and test things and wants to get deep into the
technology there's not a lot of like it's really hard to get users like that and then also have their trust if you like continuously make different changes and stuff like that um but so I really thank you for thank you thank you for all of that um I mean it's the least I can do you you guys are the ones caring here uh so I'm just a consumer at the
end but uh I really appreciate the job and how fast you are also on taking everything serious and just there's already a fix on the next release so Tuesdays are my favorite days nowadays with all the here um there's 207 of us here now welcome oh yeah daniel I have to do these recaps by the way sorry to interrupt this because I have to do a
recap yeah go for it if you're just joining us we're here with our homie Danielle from Stockholm uh he's a MRA community member and before this we were talking about Johnny IV and OpenAI's $6.4 billion acquisition i talked about acquisitions in general so if you wondering how those things work we did a
YC rapidfire we saw MCP marketplace which was not the map was not very good to be honest but we made it on there so that was great and then now we're kind of getting to know Daniel here so thank you 207 people for being here back to what we were doing so um where where do you work now Daniel so I have my own company and I
work as a consultant uh so currently I've been into different industry sectors from like uh telecommunications or fintech and right now I'm on sustainability uh working as a integration uh developer uh so uh I'm trying to put some like those agentic solutions in my team but we're not there quite there yet so uh I I still I'm I'm the kind of guy that needs to keep on learning and keep exploring things so
like I it's I've been tweeting about it once that my hobby is also my work i really enjoy building uh and especially when I can do what I want like I can be the one uh choosing my path so um yeah uh I'm currently as I said as an integration developer through my own company um but uh on coming home uh when
I get my time alone then I can hack into my projects in which I'll be talking about one of them here um question about like that hobby uh hobby is work or passion hobbies or work is passion and all that um I feel like a lot I I feel like a lot of new developers don't have that passion anymore um I know when I
started and when Marvin as well started like it was I used to love getting home and saying I don't have to work on that i get to work on my own stuff and I get to learn oh flux library came out flumix Redux I got to learn all this that was a long time ago i mean it was so overwhelming to catch up and it still is so like if you're
really passionate about it you have to stay ahead of the game uh because you at the end of the day on a daily office job you you won't get the time some some companies of course have like hackathons or have like uh you know education days but it's not enough one day or one week is not enough so you have to put that extra effort to catch up on all those tools that especially on the TypeScript
JavaScript environment is just insane changes all the time you can't catch up we got this good advice from Matt Bilman from Nellifi we went over to his office one day because and this is the beginning of MRA because we were really actually we're introducing another tool that someone will need to go and learn and put some hobby time into and so he
was like hey the reason why Netlefi was very successful in the beginning is people could build a Gatsby blog on a weekend and deploy it to Netlefi so you feel like you have some project that you can like sink your teeth into and you can make it however you'd like and then at the end you can get it in the internet you know and so
all those things are super valuable in the user journey right um so we kind of took some some advice that advice very seriously that is true I mean I remember there there was this ex extensive uh templates uh that you could just grab and uh deploy them straight away so Uh I was one of those hacking those weekly websites and we'll say with one of those templates and just tweak it to your needs and
let's uh let's get into your projects i would love to see uh what you're building yes uh let me see if I can share here all right let me add this to this there we go can you see my screen sir whoa this looks great looks nice yes so uh a bit of a background story um I started building agent like the first
lines of code were in u December right after Christmas you know the talking about the holidays you get to refresh yourself a bit more and I'm like okay I have to get my hands on this uh so I I didn't really know what I could do or build but I've I've always been very passionate of trying to uh make things easier for the nontechnical people uh we
are very technical and we maybe there's a lot of power users here on the stream that they like to uh go straight into the latest releases of cloud and try and test and so on but it's a bit of a l like it takes some time there's a bit of lag for for example my parents to start using this technology uh so I my goal
here was to try to abstract that and to try to make it easier in a more like friendly language when you can still use the technology and the primitives of like everything in the AI environment but trying to make it in a simple interface uh and at the end of the day it's the same same but different you
know uh so yeah um since March by the way I I I went out live after three months I went live with my first iteration and I was not quite satisfied because I was looking more for some agents that could grab a task and execute it and I couldn't really achieve that with the tech stack that I picked by then and that was actually when I
pivoted to master so I was like okay I need to redo my back end uh But um it was nice because the design of it was already just change the backwards compatible APIs and we'll be able to to achieve that and master exposes those API so easily through uh just like the box and it it's very u customiz customizable as well like I I wanted for
example to have a middleware to uh protect them to try to don't expose those endpoints and super easy uh so uh yeah it was a very pleasant the refactor experience and it didn't took that much time I must say so um yeah uh the purpose of agent is I wanted to make these productivity agents uh that can help you boost your well productivity
and fit your needs uh so I don't know who has let me first log in here the demo account and you can imagine that an agent is sort of like if you go to the go going back to cloud desktop they offer projects and projects is a way that you can isolate some sort of system prompt and you have also your your
conversations so you can imagine that an agent here is sort of like a projecting cloud uh but I'm trying to give it a bit more like a friendly vibe you can have you know a character as well that you can either pick the image yourself or or it will be generated by uh by the system uh so here's the interface you start your flow by selecting what kind of problem
you want to solve like what kind of agent would you need and for this demo I was thinking to maybe make an agent that uh will help us compare different products uh and analyze the key features of them so what's going on here and I really love this way of working that I have an agent that creates an agent
which is amazing so I can pick the model I can tweak the instructions of that particular agent and I can see okay this is what what I want and test what's the best results so uh yeah right now we got the agent here but now I'm giving a bit personality and what's happening here it's a master workflow uh to uh what's it called uh generate an image for it with the new open AI image
gen so uh we wait a bit for that to be baked are you doing dynamic instructions on an agent and then you're like loading like different metadata that you saved in the database sick so agent is a multi-tenant uh architecture and that means that you can have many users but they are consuming just one instance of an agent that is dynamic or in my new updates right now
it's not only one instance it's going to be many instances because I need to collaborate to become one entity right yep uh but for now uh and This goes again like the flexibility with Mastra was on how you can make your own custom stream endpoint so I made my own custom stream that it's just a an extension of the current one and I can send different
par parameters into the request such as like my user ID or like in this case this is the name that that the agent got when it was created i can send that as a context if I want as well or maybe like I have two different database right like the one for master where it stores the conversations the threads and all that and the other one that is the classic
CRUD database that will store my user metadata like okay how many agents they have you know and they have their unique identifier as well so I can use all that to send it to the stream endpoint and sort of make the logic of okay this is how the uh resource uh ID will be splitted or like how it's going to be unique for handling like unique conversations for each particular user
for example so yeah and of course each of these agents will have their own instructions that are also injected into that stream on demand uh so the interface is pretty uh simple uh you are familiar with the classic chat interface you can create new chats you can see the threads here um you have the instructions here where you can fine-tune them if you want by hand or the agent can also do it itself um
settings also you can rename it can tell the agent to I want to call you uh Lisa uh so with the master tool you're I'm able to call my home back end and make that change and then on the front end I just update the interface for UI so it looks like yeah uh as I said you can change your avatar if you want to have
your custom one um I noticed that it's starting this this uh workflow started to get too enterprisey and all of my agents started you know showing off with a big suit or like there was a one that I chain but a suit and I'm like okay cool flexing here that's cool so uh yeah one of the things that it's been a lot of talk to is MCPS of course and um I
wanted to when I started trying them like okay you can try it in curser or you can try it in like cloud desktop or similar tools like that but you have at least back then you have to jump into the settings go to a GitHub repo look all the five different ways of installing it try to understand them and wish luck to see if it works or if you just broke
your whole whole settings so I try to make it a bit more friendly on how to MTP so you have the easy way or the advanced way the advanced way is giving that flexibility to still uh okay you know what you're doing you have a um HTTP uh remote uh MCP server that you can just add and then you can set up environmental variables or you can have the easy way which is the ones that um
I've tried I've already tested and I sort of like filtered that I know they work and going to make it as easy for you talking about composio before uh I started using them as a sort of provider for them uh just to try to get try to give something already out of the box you can go to okay I cannot if I
press you cannot see my screen right no I can't see the composer okay but well they have a uh this kind of library I haven't added all of them but uh it was a bit challenging how to integrate with them because they don't really provide a SDK to achieve that like uh out of the box but for this particular demo I'm going to be uh installing a Google
Sheets so we can make the Google Sheets um uh sheet to store the result of whatever this agent is going to be comparing uh and the flow goes just like that so first it does an installation then then it does a validation what I call it which is basically uh composure generates a unique URL and it's a secret for you so I'm storing that but I have
to encrypt it and I need to uh also ping their servers to confirm that it's working so uh another thing speaking to my target group here is like non-technical people i I just didn't want to say like oh yeah success and then what so I'm trying to make it like a bit more uh easier like hey pop a message directly to the agent and start the authentication process i'm going to
press here uh you're not seeing this but it's just me accepting permissions uh and then I can just can you confirm access this is so cool man yeah it's very advanced to Thanks i'm just so happy they're using MRA too because I I kind of know I just know how you built it because obviously I Nice makes sense yeah uh one of the features I've been uh exploring now is
to give the possibility to swap models on demand so I'm I'm also trying to vision this product as a cursor for non-developers that just need it for their daily tasks and integrate uh different tools that they use every day so uh I'm going to I'm trying to work on an auto mode as well to sort of let the
agent decide what's the best choice depending on the task and uh yeah I have a lot of cool features that I I believe I'm able to do master and the master's road map as well because there's a few of them they're not quite there yet but I am I'm not I know that they're coming um so uh yeah uh the next tool that
we're going to be installing for the demo uh I need to speed it up a little bit more uh it's Exa this is like their official MCP and they just released the HTTP uh remote MCP so I'm going to add my API key and this one doesn't need the validation step so it goes a bit faster the difference and the message here is
that it's uh I want the agent to enhance its own instructions based on the new tools so I use like okay give me the tools available and based on that I have another agent that helps with that instruction let's go yeah that's very smart yeah tool descriptions aren't the only thing you need to do right exactly and I've been uh noticing that
now that I have like the multi-model or like the the user can select which mo to use uh the instructions are quite important depending on the model that you're using model that you're using so I I've been thinking to expand this to have sort of like instructions based on the actual model that you have selected to get the
best results of course yeah uh but uh for now it's just one generic solution so uh yeah we can use this model for example and I can just do a quick uh prompt here we have the two tools in place uh and I'm going to ask uh to compare Mastra versus other frameworks and make an analysis just and make a Google sheet called AI frameworks comparison and do this for me
five bucks that we lose i'm scared oh no oh no okay some demo we Yeah and this is another thing that uh I think that was one of the things I contributed i wanted to sort of show the error back to user when something went wrong with the MCP connectivity so uh that was that was nice to get um and thanks a lot for that yeah thanks for uh merching
it it made my life easier as well uh let's just try again that Google validated nice you should write a book about all the things that you've done that you've done on this project because it's so cool this dynamic stuff is so cool thank you thank you i appreciate it um I'm still learning a lot and uh there's still a lot of room for improvement but
that's the beauty of it like that's what moves me to come back and keep it up uh so let's try a new thread i'll send it again that should be in place so yeah this is also like the tools just we can see uh how they're rendered come on i already try again you know Murphy's law every time you do live demos something has to happen
check active connection nice all right cool uh okay that is why because I didn't send the message of the initial prompt of comparison here we go this for all the marbles moment the moment of the truth i I already did this and it was quite good result so don't worry guys murphy's law dude come on that's true
that's true uh so here we go we're using the the the X uh tools uh and uh yeah grabbing some information and at the end we'll get in place uh so yeah I can access it here but I will need to yeah share that on you cannot see that screen but yeah it's just you know the MCP server deleting that so cool this is so cool yeah so uh few things I've been uh
trying to improve now is you know sometimes I want to do bigger tasks want to crawl a 100 pages and that's when of course the prompt gets too big or so and I I am exploring now ways of how can I split that task into subtasks and using the power of workflows with mastra to sort of make a a validator of knowing like hey you have to validate that you
actually did the hundred uh you know links parsing or so or crawling because I've been getting a lot uh it just like gets lazy and oh I'm done after 30 yeah i'm like but why you know yeah so I think that's the the beauty of theic and workflow that it's I think I'm pretty sure it should be possible with the workflows just need to Yep get my hands on a bit more on that
yeah like you that's it yeah you pretty much want to build like an evaluation kind of loop and you just keep going until you're done exactly exactly like you have a a task and you have an agent that tells you this is the task you cannot exit the loop until you Yeah the validator allows you yeah uh but uh yeah
you Tony shared yesterday uh to the community the lab and I I was like this is exactly what I'm what I'm working on right now we are cooking so there'll be a lot to inspire for your uh your product here and hopefully you can use it to just do that as well so um like with anything man if you if we have if you have any problems or if you want to get your hands dirty in there too um let
us know also quick question i'm curious like Yeah um like what's like the goal for this project do you want to like get like it's more for introducing it to non-technical people um have you tried have they tried it yet what do they think yeah so I have a few users but I think I I haven't put any like marketing like focus right now because I I want to
reach that sort of uh state where it's like bug free i mean you can't really achieve that but I want to reach a bit more stability make it more like serverless first uh everything in the infrastructure uh but yeah the goal is to try to like do this every day yeah to keep on working and iterating on it and uh I
I I've been doing a lot of projects during my career and I think one of my the things that I've noticed that I have failed on is that I go live with it i'm good on achieving the MC MVPs but I'm bad on continue delivering because you know as a builder you have a thousand ideas in your head and you just want to jump to
the next one yeah buttal I think the difference with this project is that I I have told myself that I need to commit and it doesn't need to be that you have to stay stick to the same formula but you that you have to iterate and you have to pivot and to try to find that the target group of people that will be
using your tool for like in a daily basis for example yeah so uh that's my goal trying to find a market fit and uh do this day that's super great man yeah so cool and if I can offer some advice on how to stay consistent because we we we go through the same problems not right now cuz Monster is very young but at Gatsby we were there for many years
like six years doing the same every day that's And but it's not the same right but it's like doing the same thing and you're like "What other companies are out there should I be working in this JavaScript thing you know you're probably but here's my advice or here's the suggestion and this is something LeBron James says so I didn't really make it up
is like am I better today than I was yesterday and that's the only thing you really like is the product that I'm doing or working on is it that much better today and if I do that then I can do it every day you know exactly that's that's a good thing to to take home and I've been also thinking like just the fact that I'm working with this every
day it's just opening my ideas into like my and my knowledge as well and keeping myself relevant so it's just I think at the end of the day that's what is more valuable regardless if the product makes it or not to a market field cool uh so yeah I'm very happy to to be sharing this here and uh we keep in touch yeah we'll keep in touch daniel thank you so much for coming that was awesome thank
you for the invite yeah of course we'll see you around on the Discord byebye man that was just crazy wow that was so impressive it's so crazy to see what the users could build with your our our framework you know I was lucky with my demo on Friday and now I'm like hey I was just eye on the side you know that was just crazy man so on Fridays we do demos internally and Marvin built a
product but honestly after looking at Daniel's product dude I don't think that was a product anymore you know what I mean i'm done we have 312 people in here that is wild um let me do a recap and then let's go through some chat messages and then we are going to go to our next segment which is meet the mstras and I believe they the mras that you're going
to meet are already here um 312 people all right let me tell you what we've been talking about all right uh we started off with AI news johnny Ives company gets bought by six for 6.4 billion we kind of talked about whether that's a incestuous good bad or a very good thing for the future of AI we looked at a MCP marketplace board um
which was okay it was like an okay image um but uh we were on it so that was cool um then we talked about we did YC Rapid Fire we just went through a bunch of AI companies that I'm you know I like and uh then we had Daniel come on um and then uh Daniel showed off his uh project called Agent uh which is dope and he just gave us a demo and so that was
amazing um so without further ado we're going to bring on the next mastras uh Kenny and Dio so welcome Kenny welcome Dio what up hey guys hello how are you doing hi everyone good good also chat um let's go through some chat messages but chat uh we we forgot to share uh Daniel's uh project the agent project so if you want to go check it out and I
really recommend you go to go to agent aig.sh and go give it a try give him some feedback you know like man that was really cool but uh okay what up Kenny what up Dio how's life yeah do you know i'm just chilling chilling tell the tell the audience where you're from want to go yeah I'll go first so hi everyone i'm Kenny i'm from Lagos
Nigeria and uh yeah I work at Mastra as a design engineer so if you've used the website use the docs read the blog um I was responsible for most of those things and yeah it's been it's been pretty fun over to you D um yeah thanks Kenny like Kenny was being very humble but like he's an amazing fronted engineer like he's he's
cracked um but yeah I'm Dio i'm also based out of Lagos Nigeria and interestingly right now I'm in Kgali Rwanda just on like a very small vacation um but yeah I do a lot of so like I'm originally a front end engineer but since I started working on master I've just done so much stuff i'm not really sure I can call myself
yeah I'm not like it's weird cuz I'm not even sure what to call myself right now but but yeah I've worked on a bunch of things from workflows to some parts of evals and just so many things um right now I'm working on some design stuff alongside Kenny um making some improvements to our doc chatbot as well
maybe some of you guys have used it um in person but but yeah that's that's me you both of y'all are being too humble these dudes came to MRA from the very beginning before we we were even MRA we were working on some other and they've just been kind of the best type of engineers are the ones that will do anything and everything right doesn't really matter what it is they're down to
learn and do it so I'm glad to have you guys here i know you guys wanted to share something that you've been learning before we get into that I want to address some of these comments here in the chat so first one thanks Amin congrats to us too yeah then uh rude people die young welcome back he was he
was here on Friday i think we called him rude boy or something um we really need guru level people in our community like you guys thanks dude thanks rude boy um next Georgie are you guys in a different time zone i'm enjoying this earlier streams i'm in Europe um we discovered that EU EU live stream is tight so and I'm visiting Most right now so we've
been doing one in the EU time zone and then one in the US time zone i'm thinking we're going to keep doing this even though I'll be going back um maybe Marvin here or Ward or maybe even Kenny and Dio will be hosting these um next when swapping models on the fly is the prompt updated to bet match the practices of each model daniel did address this when he was here the answer
is yes and then I can't pronounce the name but cool cool and then lastly Rude Boy gave us some hearts so um anyway thanks for everyone for being here there's 321 of you here which is absolutely bonkers crazy honestly and we're just two guys just around so thanks for being here the bed of my baby just Yeah this is a little crib here uh we
were speaking French in the beginning there's some French people in here too so um but yeah back to the show kenny Dio I know you guys are dying to get on to to tell everybody some things you learned so take it away yeah uh I can go for it go for it so we've been working on you know writing some evals for our dogs agents our chatbots which some of you have probably used where you can like ask about things
relating to master in natural language and returns to you but then one of the things we've always wanted to do was increase the quality of the results and evals are like one of the ways of doing that and so we've been working through a lot of off-the-shelf solutions that we've got so right now Mastra has like a bunch of offthe-shelf solutions so
things for testing hallucination for testing contextual recall for testing um you know I think you know tone and all other things like that so right now you know two major ones we started with are like hallucination and so in hallucination for instance you need like a context like you need a source of truth where you can like you know be sure of the or like be sure of the data
and then run what's the agent is returning against this source of truth so I think for master we had this llm.txt txt file which at the start looks really promising but then we ran into an issue of just exhausting the context window because the lls.txt file is like a very huge file and so one of the ways of solving that was figuring out how to have like granular context
and then using that on demand for each of the evals we're testing for so it's like okay if you have your context you probably have to figure out how to break it down into chunks and then use that because if not you're either going to exhaust the context window or exhaust your token per limits which we kind of
run into when we use openi solution uh I think one other thing we discovered uh is like timing stuff we're still trying to you know you know tighten that up make it better but it does seem like you know streaming makes things a lot better in AI world and when you don't stream stuff it takes a while like for the agent to return a response so we're
trying to figure out how to probably solve that problem so that you know our evals don't take forever in our CI um but so far so good we've been having like some success and these evas have actually you know helped us realize you know some things that could be better so oh maybe our prompts could be better to
have like a better responses or maybe you know we verify that you know tools are not being called and so we actually need to do something to make sure tools are being called by the agent so yeah um pretty pretty informative we'll have more to share as well but I'll pass it over to to Dio to to give more
information from you then um yes so thank you Kenny for for the summary and just to piggy back off what Kenya has already said um we have a bunch of evals we have a bunch of offtheshelf evals and I just want to kind of highlight that like Bashra also has what we call custom evals as well so this essentially lets
you kind of custom bake an eval to you know your application needs so for instance like our doc agent uses or like leverages on what we call like our um MCP doc server and so the correctness of the information you're getting when you query our doc chart but depends on if it's getting the correct context from
the tools it's calling so there's one thing for it to hallucinate giving the correct context but then there's also the question of is it actually getting the correct context is it calling the tools it's supposed to call and you know we didn't have an offtheshelf from Eval for that so we decided to build one
using um our custom evals primitives and so this kind of showed us like okay for different query calls um our primitives say for workflows or for agents um where are the tool calls coming from like if we expect to call to come from examples or to come from docs right is it actually calling these tools and like is the agent actually getting the correct context and so that's like the first step before you then get to
hallucination so like it has the correct context it's calling the correct tools and now is it actually hallucinating so that's just kind of how we see eval it's like yeah we provide this offtheshelf solutions which depending on like your agent and what it's doing like you pick and choose which is appropriate for your needs but also like that may not be
enough right it's exhaustive but it may not be enough and so you have the opportunity to also like custom um custom bake your own thing custom roll your own thing and you know it works pretty well um our agent is like in a good condition right now most of our tests are like returning the correct results and what we just do is fine tune
here and there depending on like um any variability and so yeah who knows maybe like I mean there's TDD in the in the traditional um software engineering world but maybe we'll have something like an ED for instance or like eval driven development where like you're writing your evals you're writing your
agents you're taking your prompts and like you know you find you adjust and evals are cool but evolves are hard and Um yeah we're trying to make it better here in Master so D let's show the J let's show the code um to everybody for the docs chatbot and the evals that you were working on just so we can uh put
some visuals to it okay um how do you share a screen yeah oh hold on let me just make sure my windows are all cleaned up give me a second okay um share screen yeah okay so all right looks good so first thing I'm going to show is um I'm just going to show you hold down okay three one one more i think that's good maybe yeah that's good
all right okay so our doc agent um right now this is it's just a simple just a simple agent which leverages on what we call as I explained before our doc mcp server so to MCP server provides like a bunch of tools and we're basically just destructuring like the specific ones we need for for the chatbot and so that's
what we do here and we pass down into um into the agent as tools um you also notice like we have this extra link check out tool here and this is just so cuz sometimes like you ask a you ask the chat some questions and we've told you to provide links so like you need to do more research or you need to like
validate it responses against like where getting this info from so provide links to maybe our examples or documentation or references or things like that but sometimes like those links may not exist right where it like hallucinates some links sometimes and so we defined this link checker tool which just like
validates that like those tools are or rather that those URLs are actually correct and like linked to um to the correct resource so like every single um link that gets like passed back as a response always goes to this link checker tool and like you'll notice we provide like a bunch of like we really shout at the agent to make sure it's it
validates every URL before before sending it back and um yeah so like the agent itself is pretty simple most of the work is done by the tools we provide to it and then just like tweaking this out of the prompt um but yeah so that's the agent itself and then let me show you the custom metric we wrote for validating tool calls um so let me start
here okay so just what makes up a metric in the first place is essentially a judge right so this judge has a few things it has like a mod can pass to it and this is kind of the crux of of um creating a custom creating a custom um eval is the ability to override the judge and then also like provide custom ways for you to like measure but also like calculate the
scoring for for the evil and like you'll notice here that like for two calls it's quite simple we evaluate the judge and from evaluate here we're essentially calling a model with structured outputs right so we get um a boolean back has two calls and then the two calls that were actually that actually made for for that particular query and then we're
able to calculate score based on what we've um defined here it's simple but like depending on your use case it might be like super complex it may be based on some some complicated algorithm whatever but here it's just a very simple Did you make it so or did it not and and yeah so let me show you the judge itself and um yeah so the judge itself
just has two methods right it has evaluates and it has get reason um we're mostly concerned with evaluate right which as I said earlier um just essentially calls the agent passes it a prompt and then gets back some structured output which we then use to calculate the scoring or calculate the evaluation um but yeah like you can see
like like all of these things are like pretty well laid out where it's like you don't need to think too much about like how to structure things it's you just like inherit from a base class and then you just change what you need to and and you're good um so yeah I think that's really the crux of it um for using an
offtheshelf eval it's quite straightforward you just define it provide the provide the parameters you need and you pass it to your agent something we're working on that doesn't exist yet is being able to interact with the eval in the playground um right now they don't show up so we basically um write our evaluation tests this way right we kind of leverage v test or you
know whatever you want to use to get the evaluation to get the metric um generate a response pass it the parameters you need and then essentially just write expectations against the score um depending on like the particular evalities so so yeah right now we are basically using it in CI and in just basic test but eventually we're able to actually interact with this in the playground um change your inputs change
your outputs and we run your Evals um that way it's really really EDDD um but yeah that's really everything I wanted to show to you guys today sick that's a good one that's awesome that's awesome thank you guys you know when we were in Japan they um they kept asking me like "What's the moat in AI applications?" And I kept saying it's probably the evals because
uh if you can have confidence that your agent responds properly in x percentage of situations you'll always beat the person who has no idea you know um because and Dio correct me if I'm wrong when we released the chatbot we didn't necessarily know how it was performing right until we started getting like people saying "Yo what the is this this is incorrect." Whatever and then we were like "Oh
snap." Like we're not we we also did zero emails right until now um so what did you learn from then you saw the issues that the customer was having like how did the eval kind of reinforce your knowledge are they frozen think so oh no is it us i don't think so it's the software somehow oh Kenny's dropped out can you make a sign if you're in the chat
where'd they go uh oh hey Chad are we on or did they gone just check i go live what's happening i have access to YouTube uh oh they're both gone we're here uh oh yep we're We're still here okay here maybe they just had Dude what if all of Nigeria just went down internet went down all right we'll see what happens with them um Yeah okay thank you thank you thank you thank you thank you so you
know there's something that Daniel has said two things actually apart from the eval thing that can be interesting for people in the chat that know a bit less about the thing than we do because we are deep in the deep in the thing the first one he was saying hey I just had to extend a class and modify one method
and I'm good so it means that you're at this level of customization where if you want to do something custom in the in the framework you just extend a class modify what does not work for you and you're all good this is something I appreciate a lot about this thing is crazy oh the homies are back homies are back
welcome back we bought you Starlink and this still this happened i thought low is um hopefully Dio will come back we were saying maybe all of Nigeria just lost internet for a second yeah well like um he's in Rwanda right now so that's uh that's also interesting ah that's hilarious um I guess we'll just
you know we'll just move on we'll move on so Kenny thank you for being here dropping and coming back why don't you if you want to let people know where to follow you um and anything else you want to let them know yeah um I think um yeah I have a I have a Twitter where can I drop stuff in i I'll drop it for you okay yeah and then yeah
if if there are issues people have with like using our documentation if there's something on the website that's not clear something on the blog that's not clear just you know tag me on Discord it's Kenny there as well and then we're open to just receiving feedback and making things better so yeah thank you very
much later dude see you Kenny see you Kenny kenny's a dude he also his profile picture and Slack is Johnny Quest so that's tight um but yeah see you Kenny till next time also can you send me your Twitter handle because I can't find it yes and then I'll post it you can put it in our private chat here private chat okay yeah
yeah cool yeah all right that was that huh yeah super cool super cool now we get to build stuff huh yeah someone was just sending me a message flo was like "Hey guys can you build something?" Oh really i mean on a joking ah okay well that's what we were going to do so um okay so let's put So yeah if anyone wants to go follow Kenny that's his um Twitter
please go follow him um but yeah it's time to build stuff huh okay so now we're in the last part of the show we're going to close it out with building some stuff there are 356 people here which is wild thank you for being here um let me do a quick recap before we get into this um I hate doing these recaps dude but uh yeah this is AI Agents Hour hosted by MRA i'm Obby
this is Marvin uh we are in France right now just so happen to be here um we talked about AI news Johnny Ives company getting sold or bought sorry bought to um OpenAI hardware and AI we talked about that we saw MCP marketplace map which was you know whatever but we were on it so that was cool uh we went through a YC rapid fire which is AI
companies that we like and then we had a MRA community member Daniel show off his uh product called agent i'll just post that here and then um then we had Kenny and Dio come on uh to talk about evals um Dio's internet's back so I'll let him come on and say goodbye because that's the right thing to do dio welcome back hi guys yeah hotel Wi-Fi is
Yeah I don't want to swear but bad words um but yeah um thanks for having me Abby and Marvin um hope you guys were able to learn something about Evolves and yeah we're like constantly trying to make things better so we'd also appreciate like some feedback from you know you guys if you've used it um same with the chatbot as well if you have any feedback for us we Yeah we'll get on it as soon
as possible awesome yeah keeping awesome and enjoyable you're the man dude bye guys okay now it's time to build stuff yeah that's a team man that's a team those guys are cool yeah all right so what do we build today well okay okay so this is this comes from a conversation we were having last night which is like can we like make music with with AI just generating
and so we want to figure out if we can um and then we'll vibe code a bunch of stuff to make it so but first what we got to do is come up with like our milestones that we want to hit and we are going to speed code or we'll try to so um let me get going here to be honest we don't know yet how we'll do so it
will be experimentation experimentation experimentation exactly i'll share my screen um this is screen one i really like this part of the show when you are when you're just in the terminal you know this is a part of the show that I I love getting to um we'll create a new master project yeah maybe you can give an intro for
people that are like not super aware of MRA yet i mean some people are just joining for the first time the French the French people for example the French friends if you want a new market uh yeah so for those who don't know MRA is a uh open source TypeScript agent framework we have a nice little create
MRA experience here which I'm going to use to um bootstrap my project um and then from there we'll probably add some more stuff and uh honestly we'll fumble our way to victory um but yeah the create master CLI is really nice it um uh it yeah we'll go through it so this will be called DJ MRA i'm curious what model will be the best at music you know ask some Yeah yeah
we'll ask OpenAI if anthropic is good at at music which one is the best suit to generate music somehow let's install in depths we'll hit that um I guess I don't really I guess we can show this but I don't really want any of these um but you can add like when you when you do this um create monster experience you get like a default weather agent and like a
workflow um which you can then play around modify etc except both of us know what we're doing so we'll be fine and I don't need an example uh Dio was on and he was talking about the MCP doc server which powers like our docs chatbot experience but also can make you um you can have that same kind of knowledge base within your IDE i already have it installed globally so I don't really
need to do this and there we go you get a new project you can like cd and then mosa runs comes with the dev server there probably won't be anything in here but we can find out uh cool right so this is like our dev server we have agents networks which we'll talk more about some other time tools MCP servers workflows runtime
context uh it's dope but we don't have any agents yet because I didn't bootstrap any so okay so that's where we're at this base level of doing stuff and um let's get back into the IDE this is a bug that we need to fix but it's okay storage initialization is throwing an error for some reason it's because I have no storage right this is this is a bug that
uh it's not a bug or it's an intended unintended consequence or whatever of uh the breaking change we made last week where we make things explicit now um but I don't really care about that um we have a question real quick can you we use Playright MCP in agent or tool yes you can you can use any MCP server we have an MCP client so have
fun okay so now let's we need this oops and let's open Wind Surf i'm a wind surfer by the way ouch cursor guy right here um okay all right so now let's kind of like make like a plan yeah so I've asked the I mean GPT just to know what would work and he's saying if you want to generate music you can use Google deep mind music LM but we
don't have access to it through AI SDK mh so what we could do instead is having you know MIDI support or something like that using GP4 for example uh I like the MIDI we can do some MIDI stuff there's MIDIJ JS also midij JS oh maybe we can use that which is something that probably has samples for that so let's look at MIDIJJS
so hold on one second um I use so yeah I use Playright on Cursor Global i talk with it with the chat cool thanks for that information can we see a demo if you have the time i'm sorry we do not have the time um but maybe next time but it's also very chill you could just do this within 5 seconds yourself but we're on a mission right now trying to become DJs so you can feel the Yeah
that is on a mission right now oh yeah true when am I when I when it's coding time it's cutting time okay so we'll look at MIDIJJS let's do that let's do that right now midijjs i mean so the plan is we are going to ask for a musical keyboard for some music let's say play me the Mario theme or something like that they will give the agent will give
us back some data formated somehow and then with a tool we will just play some MIDI stuff is that really that's the idea so I've never used MIDI.js me neither is this the one uh JS app on the web musical keyboard library this one I guess yeah that's the one it's really all right man that's okay cuz what I'm going to do is I'm going to vibe code the
whole thing right now cuz the like we're not going to necessarily need to do the AI pieces yet or dude actually what if every note is a tool call can be good but it can be slow it would be very slow maybe maybe yeah get split in chunks let's first make our agent then yeah so that we can just get back some output and we don't have to be pro musician
here just to be able to type some notes I think at first to render a file too yeah i'm not sure if we want to generate an audio file maybe we do at the end but I I wanted to like I wanted to play notes yeah or put a play give me a sequence of notes that I'll play or via client side tools play the note yeah I think client side tool we should play medj in the
browser directly without yeah in the browser time so that's true we got to make moves what is we're going to do something like link flow for um I'm going to keep going we'll answer questions a little bit so we'll give you some context for people that have just joining so what we are going to build is an application that uses MRA to actually
uh play music directly in the browser so we will ask an agent to give us some music patterns and we will try to play it directly in the browser using MIDI stuff or we will figure this out we don't know how to do it to be honest that's the show so AB is currently creating an agent in MRA and as you can see it's literally five line of code so
once you have this thing I mean except if you have a big instruction prompt or whatever but it's literally five lines of code to have an agent running and that is able to answer your question directly from so what if we can so maybe we can use general knowledge here so I'll make the instructions be like users
will provide you a what like a theme a song like artist or we can start super super direct like can you give me this mother theme at piano or something like that we can just give you back the information at some point we can you know iterate users will provide you a song um and you prov provide them a Nokia 33 10 you know melody yeah
um like maybe like a like you provide them the notes they can play on a MIDI yeah output or output i think we can make the translation maybe if we are able to define the output schema ourself but that should be pretty yeah pretty rough but I mean we can probably have something small okay then I need to not share my screen and I need to get the API key so time
not too hard to do okay so yeah to give also people context why is doing this when you are having an agent agent is basically here to get some text as an input and give you text as an output right but it does not interact with the real world uh and what I mean the real world is that doesn't trigger any side effects except this conversation and so if you want it to
have an effect or to do something concrete you have to use tools so the agent will be able to call tools which are just functions that will be executed or on the server or inside your uh application in the browser and in our case we will leverage this capability to just play stuff in the browser and to
make you know just music from it so Genjent will be doing a bunch of things and when he will be done he will say hey I give you access to I need to call this specific uh tool to just play music and it will just do it uh right away so so let's see let's see what just off the bat like um I want to play the Mario
yeah that I was having this one in my man super Mario theme song are you playing Inception with me or something i'm reading your mind bro okay so it seems that we have patterns d yeah it seems that but what does it mean i mean the first three letters mean something d yeah it means something that's cool let's try a different one it's not like
we can actually we can do something we can play this one how do we play it i mean we can just We don't even have to have a CL eventually we can just send this as an output yeah it's true or it could generate um here let me open up uh if we have a tool that is just a player actually we can call the player with those nodes what we want to do is take
the sequence and then run the MIDI file on on each of them so let's get MIDIJS i think E is a node for example yeah yeah oh boy great um Let me try to find out how we can do um we need the phone application for sure but I think we can yeah yeah right the MIDI player and this one can also be used using a script I think directly so
I'll not have to install dependency yeah but um isn't there can we play it from node audio like just the audio from the terminal or something like a script i don't know if we can do that pretty sure you can i don't know if we can play music from node oh that's a good point um MIDI node values let's do that thank you Rude Boy where's Root Boy
so to use some teachers yeah yeah let's play provide them MIDI note values um so then also let's actually use some um I want to add an example to my instructions that show the mappings of musical notes e for example to its MIDI integer integer value see what it does i've sent you a node speaker parity that
could do that output PCM oh we we but we have node we have a node speaker as well like the MRA but I think it's using that under the hood by the way didn't know about this one so we can do I mean we might be able to do this all server side just for fun to keep it you know low chill and then we can just vibe code
something else let's see yeah in 20 30 minutes it probably be rough but yeah we'll figure this out okay let's see if wind surf's cooking is it cooking i'm using the new SWE model which might be wrong to do actually but uh the SU model as they would say the oh boy let's see what happened um is anyone in MIDI pro because how do I verify this but we just
accept things because that's what we do yeah we'll see i mean geek stuff already that's fine are you musicians nope i don't know about what we're doing this is why we're building this I think okay so now we have some let's just go try this again here um I want to play Super Mario cool interesting well I mean if you the pattern makes sense because if you I
mean so now we want to play the the MIDI itself right okay okay let's make a script file cuz we're just going to do this like raw first raw yeah just make it index.ts we'll just make a script right you put system out print it and I'm done what is the what were the arguments for this function in Java yeah void zero or something weird anyways let me answer
this question one second uh what would it require to build something like lang flow not the full scale lang flow but a subset I don't know it would take a lot actually and this is more important for discord so or more suited a question for discord so I would go there and ask uh we have bunch of people there that can answer this question maybe more also there's some community members there as well and I
believe we were talking to link recently so anyway um hey what's up I have a question how can we implement web search using MOSRA you can use an MCP server playride browser B browser base excess search like there's a million of them so just use one and rude boy coming in with the clutch help simplest example we need a MIDI note start then a stop the time between
start and stop is the note length we should have you dude one of these days you got to come on and just you know make moves with us but we're going to stumble bumble and fumble our way through um that's how we learn so let's grab Monstra here get agent the DJ boom and then there's some hallucination and let's look at this MIDIJS thing now yeah this one will probably run the
client I think is there like node node music or something was it speaker just before that's just the speaker though no I don't care about that i mean I do want to use a MIDI player though so if it's only or a jazz soft jazz dude wow soft jazz um that's cool yeah this one is clite mjs window audio context blah blah blah this is dope this is cool this little soft jazz
thing um tight let's see what wind surf will do if we just let it Let's go yeah yeah and then uh so Oh yeah we're in his playground nice man you should share this app somehow if it's open source or Yeah I'll be curious to see how you did that so we're going to generate you know we'll use a Super Mario yeah it's a good Mario Bros theme and then what I want in
our structured output is uh we want our MIDI notes maybe which is like an array of numbers maybe and then our musical notes I guess is array of strings and this object let's see what we get so that's something very cool that's happening on the screen he's getting an agent and trying to generate he's sending it a request a text request and then he's imposing an output uh shape so
that when he will deal with his object the object should have the expected shape which is pretty convenient you tell the agent in plain string that you explicitly want this type of object which is super nice pretty good oh yeah there's this uh god damn it there's a problem with bun right now http parser
thing but oh his app is patented or something that's cool let's see in real time man that's crazy oh wow so as you can see we have an object with the right shape that we're expecting okay nice so what I'll say is I am receiving an object with MIDI notes and musical notes how can I play this audio let's go like that let's see let's see let's see let's see context
easy MIDI package oh oh okay okay it's okay rude boy we love you dude don't No need to be sorry all right so let me just double check what is this package on the side i don't know but we're This is pure vibe code we So whenever you do vibe coding we always tell people there's two types of vibe coding one is when you know what you're doing yesterday we knew what we wanted
to do for the triage we made a triage agent today we literally have no clue what we're doing not musician even we're not even musicians so truly we're vibing in the sense that we don't know if things are going to be correct but hopefully our intuition helps us you know but I don't think so i think it's
going to be hard okay oh my god let's see let's see let's see how do I make this okay so imported easy MIDI seems to be something okay okay okay output play MIDI notes this kind of code's kind of trash but like it's okay um and look they put small pauses between the notes and stuff i think it's for the because we don't have the rhythm yet we will not be able
to match the musical notes playing notes dude what if this just works on the first try it It probably won't though but will it will it virtual device installed AC set up i need to install some stuff let me do that real quick um make sure you have a virtual MIDI device installed on Mac OS you can install IA driver from that okay let me
go figure that out real quick let me stop sharing here whoa we have 403 people here welcome that is wild people watching two guys that are you know struggling struggling a lot um what is this thing audio maybe setup okay so and then what new configuration configure speakers oh so in the audio MIDI setup app on the right choose window show MIDI studio don't know if you have this one might be
a pretty old guy um sorry show Med Studio okay okay i'm in here you're here i see driver it's like uh choose window there is window somehow okay sure then show med studio in the med studio click the choose MIDI configuration popup do you have something oh that's something I was supposed to do that let me I'll show I'll show the new configuration and then let me show the
audience what we're bumbling and fumbling through but I promised y'all we would be fumbling through this so it's not like for people joining we're trying to make music from responses well also from the terminal too yeah from the terminal too okay so what do I do i think this is your configuration right in the media
studio click the choose configuration blah blah blah you may then choose new configuration midi configuration new configuration okay and then you can put a name on it terminal terminal there go and then same right no oh then you can attach a driver i mean oh this is what you were before with the piano I
guess so it's a guy oh yeah the configuration in the two doesn't change dude what if I um what if I tell wind surf it's not going to do anything i mean I mean it's not going to help me but I don't you were putting a a piano just before i'm like "Hey Bumble fumble and Dumble that's exactly what we're doing." Yeah we're having fun
that's fine it's just we are not selling this project we're just learning stuff um what it was like if you click on a piano for example i mean the video thing oh I'm the one throwing this error though so yeah because you don't have I guess so maybe this one if you go back to you have a piano here I guess can click on the if you click on the screen no nothing happens okay custom what does
this do look at my Look at my alt text bro accessibility maybe that yeah cool i think you can connect this to now what connect it oh where's Rude Boy at help us out rude Boy let's go so this way i want to play We got Chad GBT right here that's not me i want to play music on my computer i have to set up so maybe it would have been easier if we just did this in the browser huh i don't
know same issue no I really want to make it work on the terminal because that just be cool g music from JS and I have to configure also we never even read easy MIDI the easy MIDI library this is a problem with vibe coding because we don't know i mean we never used this library before so who has to use this one is probably
Is this what it is yep usage overview a virtual devices can be created by passing this get output okay there's a pretty cool library though yeah they have some stuff here interesting output okay so I'm going to say so there no we need to create a virtual yeah so install virtual MIDI synthet Option A Garage Band option B simple fint oh man we're going to the
browsers we're going to the browser dude this um okay i'm going to make Windserve do all this work for me um yeah reag i'm having Oh let's see what if it Oh okay show MIDI double click check double click IA driver devices online cool was that it was that Was that it now let's see what it says let's see
check okay we did that how do we apply it though can't apply double click IA driver check devices online apply what is this bus yeah whatever let's do it wow when surf tell us the whole time like what we had to do that was crazy the note will play with a short duration let's run the script let's try it let's just try it let's see but then
now we'll see get outputs oh oh okay so it means that something is happening with now oh musical playing the notes i don't hear though yeah means that we are lacking something between this driver and something that sounds something this is maybe where the maybe we surface the actually yeah probably okay if you want to hear the sound oh man we got to keep reading dude
you know the problem is humans just don't read that's really what it is but okay if you want to hear the sound you'll need a software synthesizer like okay like exactly what we just learned um let's pose the question i mean is this one super complex to to install or is this what if we did this in the browser
would it be easier cuz I have an idea for the browser if we do it there you want to doing in the browser with the web audio API yes let me show you how we can modify it let's go let's go yeah all right i was going to say hey we just need an index of HTML it's fine what can I do error what okay okay yeah go ahead make a public
directory go ahead do your things do whatever you want i'm just trying to play some music man yeah simple one trying to get this music going imagine that yeah isn't the funny part though is like the AI piece of this was done in like five minutes and then we're still doing all this classic software engineering figuring out what to do okay sure browser.ts might as well
you know what they were saying that also just before he was like hey the agent is the is the fast part you know he was having this agent with a big instruction he was saying hey the thing that is actually happening is what tools are doing the tools are the the critical point basically because now we are very well
equipped for creating agents and all of these this browser file I already know is going to be a piece of i can already tell like just by the name it is also taking a long time and like I can only imagine like what's happening um but I'm gonna let it uh should I Yeah I wanted to do all the MIDI stuff but it's probably going to bumble everything else so some taking forever too so it must be either it's
either it's building a rocket ship or something or it's doing jack i can bet he's actually trying to make some you know real time creating that thing you know behind the patent or something maybe I will let's see what's happening in index.html over here bunch of bunch of can move the style so browser.ts
TS sure oh hello there oh man look at this beast of a thing let's just want to see how it goes yeah oh my god yeah sure simulate API responses god this is not what we want this is not what we want i'm curious about how to use this API in the browser just that yeah let's just I mean I guess we should just open this
um and I'm going to have Wind Surf oh wait it's still doing stuff still having fun i'm going to let it open up the file too i don't know but we can we'll pursue it i mean we'll we'll push forward like Okay reloading i've given my I've given the wheel to the to wind surf so at some point I'm going to have to
probably step in but today is not Well right now is not that okay let's see it told us to call npm rundev which Okay um why did it do this no this was God damn it why we didn't even need this okay let's let's actually code now god damn it let's get into it we don't need to build the server we also need to don't even need to run this oh my
god dude oh my god yeah this i don't need this ser wants also don't even need this MIDI stuff anymore we don't need many of this we actually don't need any of this in the browser.ts what we do need is to import master client yeah we can yeah we can basically so two things we need to fetch the API or I mean it depends if we want to use a tool or if we just want to
interpret the the response from the agent so we can call generate like the generate endpoint or we can set a tool I want to yeah we can give the we can give a client side tool that plays audio right and then um we'll just have to generate the notes it'll all work it's just it'll all work it's just I wish we had more time cuz I got to do a whole
project setup i'm not using basic ass JS some crazy shit's happening over there um I don't know what the game is doing to be honest sanding i don't know um so that's the fork in the road here i say we just try for it let's have a simple function that fetches for generate get back the MIDI nodes and we try to run the the media API just a
simple script making a fetch request and then we can try to just spin up the MIDI API and see how it goes we have a plan that's a that's a plan um that's the plan sure i mean how do we change this browser file though we can go quick and dirty we can remove browser index index sheet we go into public we create a script.js without transpolation
we start the index thing and then we we're good to go can do something like this very or maybe simple script or what if I just say "Hey Bootstrap." Oh I'm gonna I'm gonna see if being mad at it is going to change it this is such a mess make me a Vit app um instead of this you should be fine with it and be fast you shouldn't be mean cuz I only got Man
we're going to go over time today because I just want to see this through but I only got some time only got a little time left to live so yeah because we can Yeah the most basic thing would be just a script and fetch stuff and do it's like super easy stuff input and output never mind waiting for a response i was expecting you to yeah
i thought we already installed all that so Oh it's cuz we're making another folder potentially uh Rude Boy said "I would try and use Chad GBT to render an audio file given an array of MIDI notes." Yeah but that we already know that works right like that's that's not like the the goal the goal is to do something stupid um or
novel in a way we're on track we're on track but I did learn a lot about some stuff so that's another good thing about vibe coding is if you do take it seriously like it is frustrating we don't know what we're doing but eventually we're going to get something that kind of works or does something and then I'll be like hey I learned a lot
about MIDI notes also I didn't know MIDI notes were integers right cool thank you for that um so and then also didn't know there all these like JS libraries and all that so it's pretty cool um I didn't know that you had to have a setup for running on your machine and that you don't have to on the browser for example
accept all i'm just accepting everything i don't even see the code where it is anymore slash at the root i don't know what code is it updating i don't know what is it doing where is this file oh wait oh it made another project called DJ-Mra that's pretty tight so let's go find that dj dashra now we're here
it's still going to in the other in the other one nope not here this one that's funny dude man that's funny oh yeah i'll run npm rundev sure why not just to see how it goes okay so let's see if we have a server now oh wow it actually did a lot of stuff okay let's go look at what it did the DJ Mastra thing we can also pull it
up on Oh there's a UI there's a UI and everything i think you have the port in there yeah just above yeah cool cool if you press something Oh okay cool okay cool that's something at least so isn't so bad now let's take our MRA folder and just plop it in here sick and this index that's great we're just going to use this now if I go to
here I'm going to mpm install client.js also I need to double check our MRA dependencies here which I believe we're good and maybe it already made a Monstra stuff let me see if there's any new Monstra in here nope that's just the one that we brought so all right Brett so now we'll have we'll have the clientjs so then we'll go into
I don't know there's some generate generate notes generate notes which is here what is he doing just faking it hey that's cool we have the same type of we don't have the the same matching between the notes that he's trying to play but but we can we can figure it out so let's make a master client here um so master client is um a library that
allows to access your master instance using a convenient API it makes basically fetch request to your server and it's super nice to use i'm using it almost every day and super nice just going to put this on we might get some corores huh but I'll just enable cores on Yeah yeah yeah there is a course i should have I think I have a project let's see if if we have it in
depth maybe Daniel has so generate notes let's get a agent here client get agent yeah dj DJ then we can do agent generate and that this generate should come from wherever that Super Mario is so song name so generate should come from the song name so all right this be messages song name and then uh the output maybe output will be like a JSON schema so it'll be type
object properties notes sure why not let's have the same as the use but yeah yeah and then we can get our res salt and then we can set current notes as result dot object dot might be notes notes as notes yeah we can sure we can do that then you'll have Yeah uh yeah we don't have great typing here which we need to figure out but I don't care right now
so oops about to any my way to victory unexpected any no I want to use any but I guess you can't i mean as long as the death process is okay with that you can I can live with that for this project i can't because it's not going to let me um let's do notes resultobject this should be as sure i don't care whatever let's just keep going um that we don't need uh I think yeah
maybe we put this in here when the set notes I think there is something okay sure let's see if this evil works but let's log this too okay so we need to do a couple things so if we we need to add a Mostrade Dev that could be just Monster Dev storage or No that's because I What did I do here don't have any storage i don't want
any let me get rid of this it's too difficult for some reasons i think I have um I don't have mash I have my global version on MRA is not the same as um yeah so now what I need to do is mpm i mastra in here boy is so nice he wants to help us by creating a report thanks a lot man we'll figure this out thanks a lot so then we
have that and then we have wind surf playing this but where did these keys come from is this just always that has been generated uh that's pretty nice to be able to see the notes okay so then we hit play but then it probably failed right yep oh no what's this is some tracing but I would assume that we got a Coors error or something a not cors is a beer but uh
Coors error oh we got a 500 nice probably because the environment variables are not over so let me go do that real quick and we can so close dude so close but we also have to go so we might have to redo the or do this again but let me just here I'll promise we'll just do one attempt and if it doesn't work then we'll have to cut the stream because we have been here for a while 450 people yeah that's
nice marvin give them a a recap so yeah we're just messing up right now so we've been through uh today we've been through AI news uh we've talked about OpenAI and how they have acquired a great company so I I will mess up the name for 6.4 uh billion which is solutely crazy and for hardware engineering and these kind of things this is something that I personally uh don't like much but it's
very personal uh then we've been having a great demo from a member of the community uh I don't know if we have the link yet but we'll send it again which is super nice oh snap holy i got a chair with you guys sorry i was Dude okay okay okay okay okay sorry i was just around while I was talking my bad my bad but we got to we got to bring this home now um Okay super simple
okay yeah refresh the page just to make sure yeah yeah yeah yeah so we're fine application okay so I like I said or like we did right i set up everything the server is running i'm outputting the MIDI notes i sending it into this vibe coded UI let's go let's see work man work oops the stop button worked but okay let's do one more theme like um Star
Wars yeah I didn't see that see yeah whoa okay one last one one last one yo dude that's sick because I thought we were going to fail but at least we got something vibe coding for the win dude um okay what's another one what's another What's a Oh Rick Ashley oh this one can be good oh I didn't passing yesterday evening when we Yeah yeah we got Rick Rolled at your own house
i don't know that one he doesn't know yeah maybe he will know about which one titanic titanic it's taking notes for some reason but Super Mario Bros is working yeah what's like something iconic i just want to Oh Jurassic Park jurassic Park is good we're going to give you Oh what if I did Rick Ashley never gonna give you up cuz I was giving them an artist like what if it
doesn't know oh yeah it's not precision enough yeah but it also needs more of notes yeah yeah let's try one last thing and then we got to go Jurassic Park theme song we need more notes we need more levels of notes the scale dude i think it doesn't know so this is where you got Now it's like back to the AI part we need to make
I'm on 40 on four okay but we could try different models to see like what kind of things there are or maybe we need more range in like the available notes but and also timing because Yeah so fast but we can definitely do something like ancient work super nice that was fun that was super nice let's just do one last Mario and we can get out of here
it's like looping it's not the thing there's so many things we can make better here for sure okay so let's do a final recap and we are out of here um and I'm very surprised we even got that far so that's dope man it's pretty sick um yeah so we had AI news uh talked about a bunch of stuff i had Daniel from
uh agent AIG- NT.sh we had some monsters come on board teach us about evals and how we make the docs chatbot better then we struggle bust this whole freaking thing on on how to build a MIDI player where the AI agent will generate notes um and um yeah so that is that for today thank you so much for joining us all 463 of you we will see you next time peace



