MCP.Run, Smithery, Build a startup research agent, Principles of Building AI Agents Book

hey everybody we are coming at you live Well I guess I say we Normally Ob's here with me Today you just get me Ob's on a jet plane uh heading to Japan So I'm sure we will hear from him and he'll probably be doing some live streams of his own while he's in Japan with Tony who's one of the other uh founding engineers of MRA So today what do we got

on the docket well we're going to talk a little bit about the master.build hackathon do a little update We're going to have three guests come on today The first is going to be someone from MCP.Run So we're going to talk a little bit about MCP Why of course if we're talking AI agents we should also talk

MCP We're going to talk to my co-founder Sam who wrote the book on AI agents We'll talk a little bit about how he wrote the book why and it's back here on my shelf right there I'll go grab it in a little bit and we will uh we'll talk through it And then we also have a guest coming on from Smithery So that's what

we're doing today Thank you for tuning in And first just a little thought uh for everyone out there I want to share a recent tweet that I just read for the first time and I'll give you some of my hot takes on it but happy to hear yours in the chat So whether you're watching on YouTube LinkedIn X Twitter whatever we're calling it these days please uh

comment on what you think and we'll we'll talk about it All right so let's talk about this tweet here from Matt Poc Hot take I found that AI is better on large mature code bases than toy projects Depends on good context management of course but being able to point AI at existing implementations is a lot easier than starting from scratch Huh I'd say this is kind of

nuanced in in my opinion I don't know that I completely agree but it's something we should talk about I think that in general if starting from scratch you can get further faster Now that being said that's really good for prototypes but it often breaks down where if you are actually working on something pretty

complex you you're kind of building the patterns as you go So I can really see the the value in having a large scale codebase to point at especially for specific examples But I do think there's a lot of nuance here The biggest thing that I've found success when dealing with a larger codebase specifically you know monor repos is you do have to because there is so much context you do have to guide it

You have to know the codebase enough to guide it on where it's it should actually be looking And if you do that then you you give it specific code examples you point to the specific parts in the file where it should be looking then it can do much better because it does have the examples to build on However you know if you don't if you don't know the codebase that well and

you just expect AI to uh to know the codebase I think that's where you start to run into problems So definitely useful Curious what you all think in your chat or in the chat Please send a response and I will uh I'll share it if it's interesting But we are gonna have a guest come on here in a few minutes Uh we're gonna have a guest from MCP Run and we will be

talking about some cool things uh working with MCP Run working with MRA and we will see see what we can build We always want to uh in in these things we always want to spend a lot of time building and so not just me talking or Obby talking when he's here but also uh actually looking at code and learning a little bit more of how do we build and

deploy AI agents So before we actually bring on this guest let's talk a little bit about what is master.build So master.build is a hackathon that's going on right now The goal is to build some really cool AI agents If you watched the live stream yesterday you got to see uh one example of an agent that really kind of takes

your uh I guess your logs from Chrome and can do interesting things help you actually improve your yeah just your websites in general And go back and watch yesterday's live stream if you're curious on what people are building in the hackathon Go to master.build There's still time There's this the hackathon goes till Friday So you can still if you've been working on

an agent or you're considering working on an agent there's still some time to submit something cool Everyone that submits something at the end of the week gets prizes We've had you know I think over 300 people sign up for the hackathon We'll see how many submissions we get but we're expecting a lot but we are giving out the the book that's in

the background here Everyone gets that book We're going to talk to Sam later about that but you'll get that sent to you And there's a whole bunch of other prizes Lots lots of prizes on on the line for uh people that even just submit You get entered into some raffles and we'll do some cool things there All right So let's talk a little bit about other things going on in

Maestro I had I had an interesting call today with a customer and they're kind of building an AI powered video editor It's really cool I'm not going to share it because I don't know how what stage they're at and if they want to share that yet but it does uh it is really interesting for a lot of reasons because one of the things that came up in this

call is something that I've been hearing uh really often and that is it's still really confusing on how to connect agents that are kind of set up as a as a back-end service right to various frontends and often times These are Nex.js sites you know candidly a lot of the people we talk to but could be websites web you know web is a very popular use case for frontends but we've

also had people do desktop apps mobile apps and the interactions the user interactions of how humans interact with these agents They're very important and I think people still really struggle with them Uh Jack had a question Could someone share the Zoom link uh there is no Zoom link for this Jack We're just live streaming So if you do send uh chat

messages though if you send a comment we will get them So please uh if you have questions along the way ask away and we'll pull them up and answer them as we go But I do think this uh problem back back to the conversation I was having earlier of how do you connect the front end of your application to an agent in

the back end seamlessly is a really challenging problem oftentimes you you think if it's just a chat interface well you just stream the result But this gets more complicated if you start to add tool calls Your agent's actually going out and doing things You need to surface information and you should it kind of

depends how much information you want to disclose to your user about what your agents doing But some kind of surfacing of that information helps This gets really complicated when you start dealing with voice agents because if your agent has to go look up some information well you don't want the person that's talking to this agent on

the phone or you know through some kind of voice interface to be waiting And so I think this whole user experience of working with agents needs a lot more uh needs to be improved a lot more And I do think that you know tools like copilot kit assistant UI you know there's some some tools kind of built into AI SDK

they they get us part of the way there but I'm kind of interested in in trying to both me and others on on the team are spending time thinking how do we make this even easier for people especially if you're building with MRA but just building agents in general So can I share the book name hold on So anyone that submits will get the principles of building AI agents book written by my co-founder Sam He's coming

on in a little bit So if you want to learn more about the book you can stick around And with that we have Steve from MCP Run So I'm going to bring Steve on here and we're going to talk about uh what MCP Run is but also maybe what we're going to be building today So give me a second here to pull him up And you're in Steve Hello everybody How

are we doing today good Good It's good to see you Thanks for joining the live stream Of course Excited to build something Yeah So can you tell me a little bit about what you know what did you have ideas for what you wanted to build today and maybe before that though just what is MCP run yeah MCP.Run run I mean

primarily is kind of a tool calling platform enriching agents or AI apps with all kinds of tools that um leverage the model context protocol So everything uh available on MCP run is MCP compliant uh if the name didn't give it away Um and we also have like a registry um that you could publish your own MCP servers too But beyond that we let you actually install MCP servers from anywhere on the

internet So you could take a remote MCP that is maybe Century's native MCP server that they host on their uh their own cloud infrastructure and install it into what we call a profile uh on MCPR run that lets you basically collect and configure MCP servers and bundle them into a single MCP server So it's really easy to distribute that out to an MCP client like an AI agent or a chat

application like cloud desktop And in doing so uh you're able to basically just configure a tool or a suite of tools one time like log into Sentry or provide the API keys to fire crawl and then manage that profile as the MCP server that gets attached to an agent And we'll probably demonstrate that kind of a deal uh today

Yeah Yeah Good good background there Uh for those of you that are just joining us I know we've had a lot join since I started this stream a while back Uh 20 minutes ago 10 minutes ago at this point We are talking with Steve from mcp.run We previously talked about the master build hackathon We have my co-founder Sam coming on to talk about

this AI agents book And we also have another guest coming on from Smithery later So what do we want to what do we want to do today Steve what what can we show people what can we dive into we always like on on these calls to show code right like dig into something that we can build together that we can uh

hopefully hopefully others can learn from as I am learning alongside you you know and we'll we'll build something cool Sure I was thinking we could just spin up a simple MRA agent and um maybe explore some of the Maestra features add a uh SSC MCP uh server and have it make some tool calls that kind of connect a few systems and do some interesting things that you know MCP run uniquely

provides uh and just kind of see how that integration goes Sounds dangerous uh but sounds fun you know We we do it live here you know Exactly See how it goes Yeah Um so I'm happy to drive or you can drive Yeah I mean it'd be great great if you can drive and I'm happy to give commentary along the way and ask questions That sounds great Uh

let me bring up a screen Yeah you should like go through like bare bones starting with nothing running in it and all that kind of stuff Yeah I think it's useful to uh do we have a do we have a goal of what we might want this agent to do sometimes it's useful to set a set set a goal and maybe we won't accomplish it in the time that we have but we can uh get

close or at least I think we'll be able to I think we'll accomplish it I I have a goal which is um you know we sell a lot of our solution to like AI agent companies and people who are integrating tools into their startups One way we find some of those customers is by like searching for latest AI startup funding rounds And so we could build an agent

that helps do that research find on the internet uh announcements or news about startups that have you know raised new rounds of funding and are growing Um and track their names maybe the funding amounts and their progress in a Excel spreadsheet Um and then do something with that data Maybe like just for fun we can add another sheet that like puts

plots those things on a chart just to kind of like visually see maybe over time the announcements and maybe like the levels of funding So I'm not sure if that'll all work flawlessly but um yeah I mean you know we let's see what we can get done and also as we hopefully find some unique startups and then if anyone

knows anyone at these startups you can just get them to come on the live stream if they you know if you're building something cool So it's it's you know you're helping me find future guests Steve So I like it Let's build that Let's build an agent that that I can you know take and use that to to find who we should be talking to who's who's doing cool things in the AI agent world All

right I love it Um all right I'm gonna share my whole screen Uh so apologies if you get the little infinite loop here for a second Yeah Um coach me through I'm going to do Mastra Actually is it NPX how do I get started yeah if you do uh so do you can do mp mpm create master at latest And if you can go ahead and zoom just a little bit Give me know three at least three three or four clicks Maybe better

or more two more Two more Boom All right That that looks that should be better for everybody All right Let's do it Let's call this master build demo Um while this is running maybe I'll grab a API key Um and I'll just temporarily stop sharing my screen to do so And we did get a question from Jack in the YouTube comments Is there an open

source UI that we could use to connect internal MCPS as uh I missed the last half of that I don't know if my audio cut out or if I got disconnected Yeah sorry about that Um we Yeah so Jack asked a question Is there an open- source UI that we could use to connect internal MCPS as well as external MCP so I think um I think there are as far as

like if you're talking about front-end UI uh tools there are quite a few things that you could start to look at Copilot kit assistant UI Um those would be areas to maybe look at I don't know if you Steve if you have any other comments on on how you would think about that I think I think there's a a lot of ways

you could answer that question Maybe some clarification is needed Yeah I don't I don't fully understand the question but that's just me Yeah I I would say ultimately what often happens is you connect an AI agent to uh an MCP which could be if you're saying internal as in like a you download the MCP there are packages that you know you download and run the MCP locally or you could

also external MCP using you know like server send events or you know HTTP requests to the an external MCP and typically you can connect your agents to that and then I don't know if there are some front-end tools I do I know Copilot kit's doing more things to show like tool calling in the UI but I think

there's a there's a that's a loaded question Jack so I don't know if we fully can answer it feel free to add some more clarification yeah in the meantime um I have kind of just the core master agent uh with a little example set up this weather agent um why don't we just edit the weather agent and make it kind of our uh research and call it our like startup

research ajax Yeah The the the most common thing I do in all my when I'm ever I'm getting started is I just take the weather agent and start from there So this is this is my workflow as well It's just it's a great it's a great starting point Yeah I mean if you can I sometimes remembering syntax is hard So you know agreed Um

don't need memory for this guy Uh at least we could add it later but for right now I don't believe Yeah Keep keep it simple Keep it simple So we more likely to work then right exactly What I do need though is our MCP client config right yeah Did you add the you install yeah So let's install MRA or at

mastramcp And then that should install that And then yeah there should be an MCP client Yeah there we go All right So I think we can do it out here Let's call this like client Uh is this like a Yeah it's a class Yeah you need to do new client And then there are some options I think it you need like the it's like a server

It's an object with server and then you name the server and then you can pass in like the URL But I'll pull up the docs so we can Oh yeah servers Uh then yeah I think you're right It's like the name of it Maybe we'll call this like mcp.run And this takes like a URL I think Yeah Um so I'm going to just pull

this in via let's call this like MCP run SSC URL Uh what don't you like about this oh oh string Yeah fine I like it Types helpful I think we add we add like tools here Is that right with um Yep And you And then we would want uh would want to await MCP client.get tools That's right Cool All right So we don't have these

tools set up yet but like maybe let's do that first and then I'll be able to get my SSC URL after I create my profile which is the collection of the tools I want to then give the agent So I'm going to switch gears here and go on to mcp.run um everything around um tool installation or MCP server management is

kind of condensed into this idea of profiles and a profile is like I mentioned a place to install and configure your MCP servers So let's create a new one here and just call it master demo Um it has nothing in here installed by default but we will take advantage of at least one of the built-ins that we provide Um we offer uh free access for

web research So kind of like building your own deep research agents You can just flip on the switch and now web research is here And then we also wanted to add Excel So I can install Excel into my master demo profile And when I do this it's going to bounce me out do an OOTH flow and install Excel into this profile So now I've got Excel installed into master

demo I was already logged in to Excel 365 cloud So that was kind of like a you know behind the scenes situation but you would otherwise just click authorize MCP.Run and it would bounce you back just like your typical ju just a standard standard OOTH Yep So now we've done is actually click stored Excel OOTH access credentials into this profile

securely so that we can then take this single profile and provide access to it through SSE and of course outside of MRA forgive me but we also would support these SSE URLs inside of any MCP client So goose is popular Windsor popular cursor cot cloud desktop Eventually many many things will be MCP clients and you can use these tools portably anywhere Um but

because master's has done such a great job about and adding a um a MCP client we're just going to be able to copy a URL here which is a secret effectively because it does carry some u some um like profile and and password information Um and I'm going to paste it inside my m development Okay let me um actually stop sharing real quick and do this And

while you do that for those you know 120 plus of you that are now watching we are talking with Steve from MCP Run We're trying to build a web research agent with Maestra and MCP Run to basically go out and keep track of when some of these new AI agent startups are getting funded And the goal being you know just to learn about uh these these companies so Steve can reach out

talk to them and we can get them on this uh live stream Among other among other things among just like building a cool example Exactly Exactly Um Okay So I've provided the SSC URL Um I just copied that straight out of my MCP.Ry profile here Uh just to recap again we have Excel and Web Research which is a

builtin We also have another built-in which is really helpful Maybe we could take advantage of this in a um quick edit if we get far enough along which is that we also provide you with free email sending So now you can give your agent the capability to just send email to yourself or to anybody else with an email address Um you do take ownership

of sending that email So it's like from you um just avoid spam but uh kind of a neat thing So we'll try that Um what I also like to do inside of these agents uh if I have a place to kind of put a system prompt like these instructions is say like today's date is um I think I can just to string this uh just to make sure it's kind of like fetching you know actually truly recent information Um and I'm gonna say use

your tools to carry out the user tasks as they request them Uh and then let's just run this and maybe we can use it in um the MRA UI to see if things are working Um yeah it's good to test Yeah I always find it's nice to test early that way you know Exactly Yeah And we we do have a you know we do we do have a tradition here If it works the first time everyone gets to take a drink of whatever's near them And for me it's an

energy drink but whatever it is you know so it almost never pretend Well you will have to pretend for those listening you know if if it works the first time that's a success It rarely does so it's worth celebrating right all right So how do I run this thing into the um uh uh Yeah you can just run npm rundev and it should uh hopefully just fire up And you

did copy in I'm assuming or not Well everyone doesn't have we don't have to take a drink It didn't work It didn't seem missing export Oh do I need Oh you know what i think I didn't change this here because a research agent Well that's why that's why it's the mythical thing that never happens There's always there's always a some

small thing that uh seems to be missing Yeah that was user error That that was that was my Yeah it's almost always user air Steve It really is It's funny enough when I'm in charge Um okay cool So what do we have here we have our agent We clicked into the UI for the startup research agent which was produced from the code

And we have access to these tools which is great Um it's showing us that we've got our web research tool It's showing us that we have Microsoft Excel and a bunch of the tools within the Excel server Uh and those are again all just provided through that single URL through this combined kind of virtual MCP server which is built up in the profile by

installing other MCP servers straight into it So now let's like say what we want to do with this thing Um let's say find some recent uh AI startup fundraising um news and let's say like track the company name By the way this is probably too small Sorry for that Company name Um that looks better the amount of funding raised the source article if we can

capture that and maybe say like one other thing like um and find the website URL of say like this info um or it's in a new spreadsheet and let's say called startups and track each of uh let's see track this info in rows Um I have to add one more tool which is actually the one drive tool um which is the way that Microsoft 365 like actually creates new files inside of Microsoft

cloud So let me just install this thing really quickly as easy as clicking a few buttons And then I'm gonna copy this prompt Yeah copy And we should have one drive somewhere Where where where maybe So if you refresh oh you know what we got to restart the server Yeah I was going to say if you refresh and doesn't

show up because it it probably invokes and caches the URL which will call one time Yeah exactly for for a lot of things if you were you know there are certain things you can do where you don't have to restart but adding new tools like that you typically are going to need to restart and there it is there it is cool uh we actually

also send based on the MCP spec the tool list updated notification and so I think it's like just a matter of like the SDK you guys are using supporting that but eventually we should be able to do this without a refresh because we send down the notification from the server from our server that says "Hey there's new

tools here every time you add something new to the profile." So and so we also got some comments So one by the way is this really live like can you see the comments yes we can see the comments We will we will respond And also MCP support is really nice And one more uh question while we're we're taking a quick uh inter intermediary break What is SSE could you explain what SSE of

course So MCP model context protocol um specifies a few different like official transports It is transport agnostic in the sense that I could talk between a server and a client um across anything It could be across the internet I could do an in-memory transport I could you know write literally on a piece of paper

and hand it off to my friend and they could read that piece of paper and invoke a tool call and then write me some paperback uh you wouldn't do that of course but SSE is one of the official specified transports and it stands for server sentent events and it's just a way to create kind of this um longlasting uh over HTTP connection that allows a server uh to stream back

information to a client including things like tool call results and data uh from an MCP server Um there's also the HTTP streamable version of this transport And then initially MCP released the standard IO version of the transport which lets lets you talk to tools locally on your machine Uh but things are trending in

the direction of moving all remote uh so that cloud systems can offer MCP servers to clients as well Um it looked like there was a problem with one of the tool names which is surprising Um so I'm just gonna remove Yeah search and see if that one is the one that has issue and we can add something else to

it And yeah and there's also we should be able to if we await the tools we could probably update you know we could probably a get the tools update what tools we pass in and then so worst case we can probably uh probably find a way to get that to work But let's just see if this now runs I don't expect the search to run Hm So it was not that

one I wonder if this is related to We did have an issue with the string names but I don't think it's the same issue Huh That's what if we just uh just to really isolate it What if we in our code we just grab one of the tools that are I I wonder if it's because there's like a a period in the tool name

it was MCP dot so I wonder that yeah let's try that and see and then and then let's yeah let's refresh now let's try this thing it's not the proper name for this model Yeah So we got to figure out what that They should give you type completion I think It does Yeah All right There we go All right We're get We are getting there It definitely didn't work the first time

though No one's taking any drinks Yeah So no one is uh celebrating with us It did not work the first time Okay this looks like it's starting All right So we know it's not going to do any searching because I turned off the web research tool Let me just turn that back on We'll go through this once more Luckily the tools are fast and the dev loop is not so long anymore

That's way too big Okay now we should be able to try this and see it work Okay let's see what happens Cool We call it our research tool So it's doing it's doing some research It's doing something something for us We are further along than we uh previously were While while this is running for those of you that have joined recently

we're chatting with Steve from MCPrun and we're trying to build a startup research agent So we're writing the we're building the agent in Maestro We're using MCP Run for a whole bunch of tools to go out search the web add it to an Excel spreadsheet and yeah we're going to we're going to see what we can get done in the next I don't know

probably 10 15 more minutes All right So once this is done we should have a file in here I'm just getting this all open and ready Oops And even even better maybe we can just put this up on GitHub afterwards and share it out So if anyone Sure Yeah If anyone wants to see what we build you know wants to down you know

expand upon it help us uh help us get more uh startups into into our loop of people we want to talk to here on this live stream So yeah let's just recap what it's doing We asked it to find recent startup fundraising news track the companies their funding source article and the website of the company and then store it in a new Excel sheet called startups to track this information and it has done

some research and it has found a bunch of information about AI startups It went to go and attempt to create a new spreadsheet called startups and it looks like it did do that It realized that there is no worksheet in order to like add the information to So it is trying to add the worksheet but it does not seem to be doing a good job with that It's realized it made a problem

Made an error trying to call the tool and it's recalling it It thinks there is a worksheet in there but let's check the file Okay so we do have a startup file At least that worked We got some yet So it's still working on something but it looks like it's done Doesn't it look like uh Yeah So I think you know what often happens here is I wonder if it ran out of steps So if you go in the UI because if you look there's

like what five or six there's probably five to tool calls Yeah we count five And so I'm pretty sure the default might the max steps get set to five which means that it probably you know you can basically tell it how long you want it to run You can give it a really high number in the if you scroll down uh or maybe it's in uh oh

the extended settings underneath the model settings is it's kind of hidden on the right right sidebar there right above right above your tools Yep Scroll down like four Oh yeah Yeah Change the max steps to 50 Yeah Yeah Let's just let that thing run Let it rip Yeah We we we'll burn some tokens We don't care

That's okay All right let's try it again He's doing his research We'll gladly ch uh you know exchange tokens for time right that's exactly that's the that's the what's the exchange rate of tokens for for your time All right So hopefully now since we gave it some more steps hopefully it'll do you know it'll loop longer and uh hopefully figure out its its errors here

It's kind of funny to watch these things try to use like nonoptimized software um and they just figure it out eventually Well I I mean the funny thing is when you think about it though what about when you try to use nonoptimized software you know you kind of have to click around and figure something out you know Yeah So it is

Yeah So I mean I I don't know If I if I hadn't used Excel that much I'd probably have to stumble around quite a bit too You know I actually do stumble around Office like just the Copilot 365 platform like regularly So the fact that it ever even gets uh any anywhere is is great Yeah Okay cool So it found the existing file that it already created in the past

run and it found the worksheet that it created and now it's trying to find the rest of the information that I asked it for which was go find the website URLs for those companies that it sourced in its initial research run So now it's doing more research about each of the companies that it found Let's take a look at which one of them Uh UniSoftware for financial planning AI It found

Pyperbots in Bengaloo Cool Indian company They found Toune Sutra a website URL web tune app from Google's AI Futures Fund Vigle are these real hey if you know someone at Vigle it's a meme generation platform Let's get them on the show Let's This This seems like pure a great content So Vigle where are you at you need to come on We will Let's

see Let's see what kind of memes we can generate So it did make a call Let's see if it put anything in here I'm gonna do my own research and see if these are real So let's let's see We Okay Uni Software is in is in stealth mode so we can't find their their website but Hyperbots that's real Cool Tunra

legit Oh that was neat too It did an auto resize to make the data a little more presentable inside of the worksheet Big one Look at that Very cool All right Oh and it applied some formatting Okay that's very cool So yeah job done Nice work bot Good bot Um that that was pretty cool though I mean even how far

we got in what half hour yeah less if we cut out the the chatter and the the banter and the setup Yeah If yeah if we were just you know if you're sitting alone you could have done that in 10 minutes probably Totally Um yeah we could try that one additional step which is like see if we can have it make a chart based on what it found Just like

let it do some visualization Yeah we we have like we have like five or 10 more minutes All right let's just see what it does Um let's say use this data from the spreadsheet to create a chart for the data Uh I just I'm not going to give more information than that I'm just going to see what it can do Um oh come on

Let's say use the data from the startups spreadsheet I bet you if I didn't delete the memory feature uh it would have stored the ID of the file Yeah Probably the name of the file Yeah that's a good that's a good call Yeah I bet if we and and if you wanted to have follow-up questions about it yeah if you would if we would to read memory to our master agent it definitely

would have been able to likely not have to do as many tool calls right yeah Could have just used that context Cool This is another kind of neat thing I always appreciate these little you know nuanced details where it found two sheets in the spreadsheet one called sheet one which is just like a default and one called AI startups and it looked up both the data

in them and realized sheet one was empty so I must mean use the AI startup sheet for generating the chart so it added a new worksheet to put the chart in it made a mistake and is now realizing the mistake and creating a chart data worksheet and And it is updating the data in chart data hopefully to put a

chart It's now going to create the column to put the visualization in and add the chart I don't know how to make a chart in Excel Like I will be transparent about this So it's pretty nice that AI can help me here And it's running auto fit again to make it neatly positioned And look at that Now we have a chart data Cool And it shows that UniS

software got 4.4 million Hyperbots got 6.5 The other two were not disclosed And that is correct And it even like normalized the data a little bit Sweet All right This is pretty cool I I I will never make a a chart in Excel again Well we did it We built an agent that helped us do research and plot that data for us in Excel So yeah Yeah Can we can

we you want to upload that to GitHub make it public We can we share that out for for those of you listening or just tuning in We just built a AI research agent that will go research funding rounds for startups We did it in about 30 minutes but you know a lot of banter you know some messages from the chat

filled in there A few uh human errors along the way as as you as per usual but we did it That was pretty cool Totally Uh one one question I want to answer here by by Jack that it seems like MCP is meant to consume more tokens is actually the inverse right an MCP server helps you minimize the amount of tokens the model consumes because the MCP server is running in like compute land

and it's not using GPU It's not using the model The model suggests that a tool is called and when the tool's called then that's done completely outside of the context of the model of the LM What does consume tokens is feeding tool call results back into the model which helps the model then carry out the task you act asked for it to begin with So it

just depends on what kind of data is coming back from a tool and like how many steps that the agent takes calling those tools and feeding that data back So there's plenty of ways to kind of tune the dial in the direction of reducing the amount of data or filtering it in the MCB server side before you feed that data back to the agent or

model Um but really MCP servers help to reduce the amount of tokens used because you would otherwise have to generate code using the model uh that consumes tokens where that code would then need to be executed somewhere to do like a web request to some API and it would have to know how to use that API So uh

MCP servers really actually help I think reduce the amount of tokens consumed if the agent is doing something useful Yeah And we do have one other question that I don't even know if I fully know the answer to but I will do my best if maybe Steve if you know but at what point does Ma hit MCP run to load the available tool lists so in dev what we found out

is when you run the dev server it seems to hit the URL get the available tools I uh I guess I would assume that when you deploy it it's it's kind of the same thing like when that function's invoked it should go out and get the tool It should load it in but oh yeah so it's it's right here right when we we

instantiate the agent and we await this get tools call then the MCP client makes a request to the server based on wherever this SSC URL or the HTTP endpoint is located and we await the response and the first response is the result of the protocols uh tool list request which is just the JSON schema of all of the data that represents the

tools their names the parameters they expect and the tool descriptions that describe what the tool is supposed to do and how to use it In the protocol there are other steps that the servers can and should uh um implement which notify a client of new tools made available to them This is all just so new that most clients don't have the support to handle

the tool update list notification But it is possible and we do as a server send that notification down to a client So as more of the clients mature and implement more of the spec then you won't have to restart the agent It can just get new tools delivered to it by updates from the server Uh which is which is pretty

cool That actually like brings up a really good point I think at some point in the future we need to have a a live stream where we just talk about the MCP spec in more detail because I I know I know enough but I I know there are a lot of things even at MSRA we don't currently implement in the spec and most

MCP clients don't right they are not fully spec compliant actually I don't know if any of them are um I don't know of a single open source one that is fully compliant well I I yeah I think even you know you know my experience with even cloud desktop there are things where you they're they're not necessarily always

working the way you'd expect for you know someone who built the spec right so this is no it's a moving target So I don't I don't even blame them because we we're all trying to chase the same uh the same target and building things as things are still changing But it would be a useful uh I think time to to go into a deep dive on what what are the

other things in the spec that people aren't implementing yet or really haven't even thought about I know you there are like prompts and resources and just a whole bunch of things that we're starting to implement but we haven't even got there yet Totally 100% Well let me know if you do that stream Would love to uh jump on bring some more knowledgeable people from from my team

Uh yeah happy to happy Yeah whoever whoever you think is the is your your expert on the spec and maybe I'll we'll bring some some more people on the master team that know more about the MCP spec than I do But that would be a fun future conversation So we should do that again Excellent Well thank you very much for having me Nice to chat with

everybody here uh live Um go check out mcp.run I'm hanging out in the Discord all week for master.build So please reach out to me if you have any uh questions or need some help onboarding Um the tools are free Go go go go go use them and enrich your your agents with access to the outside world All right Steve Thanks for coming on

All right And with that uh for those of you that are just joining us we have been talking about the master.build hackathon Go to master.build to learn more We had a guest from Steve from MCP Run We just built an agent that does startup research or startup funding research It could build us a spreadsheet and it uh it created an Excel chart

which is something that I know now realize I no longer have to do which is great Uh also Steve if you're still listening uh please send me that uh GitHub repo When we make that public I will share that hopefully on the live stream before we end So if anyone wants to download the code and see how it works and play around with it you all

can For those of you in the chat if you have questions along the way whether you're on LinkedIn or you're on YouTube or you're on um X just send a comment We'll see them We can answer can't answer every question we get of course but we try to answer a lot of them And with that I'm going to bring up our next guest So this uh this guest I know

pretty well Uh so this is my co-founder Sam from MRA Let's bring him on Hey guys Sam Okay Good to see you Shane Yeah Are are you at the Palace of the Dog right now i am at the Palace of the Dog I have I've arrived at the Palace of the Dog here which is you know our our I'll give everybody a tour of the Palace of the

Dog you're seeing it from a different angle and and we don't we don't we don't have our Palace of the Dog uh our Palace of the Dog poster up right now I'm just sitting in this chair which has become notorious for just being the chair that I sit in It's also a large chair I just think better when I sit in a recliner for some reason that like I don't really

understand But uh yeah hey if it works But yeah you're at you're at MRA HQ So you know Abby's on Ob's on a jet plane you're at MRA HQ and I'm here sitting in Sou Falls South Dakota But but I'm excited today to talk a little bit about the book you wrote while in YC You know not only is YC hectic but you also decided hey we should write a book at that same time So for those of you tuning in uh Principles

of Building AI agents actually if you want a copy go to master.ai/book You can get a digital copy If you want a physical copy it's very simple Just uh buy it on Amazon or uh better yet go to master.build and sign up for the hackathon and submit something you know submit some kind of agent that you built with MRA We will send you a book Exactly You got I'm glad you have your copy too you know everyone

we we have a whole a couple boxes of books and uh we we had to make it an emergency pit stop to um on the way to Japan uh like so that uh we could bring 150 books to Japan Uh so we I had to have a friend make an emergency pit stop here last night to pick up another 150 books and get them on He's just like

he's just like paying for luggage and it's just books Yeah it's just books It's books is shipping 150 uh uh books in a suitcase to across the Pacific Ocean right now as we speak It's funny We also had someone who came to the apartment and got you know probably I don't know 50 or so books and just shoved him in a suitcase to bring him to a hackathon And what yeah He tweeted out he tweeted out Georgia Tech um like like

he he tweeted out like feeling like Pablo Escobar you know with the suitcase of books which was we we appreciated that Yeah So yeah if you uh if you are running hackathons that for building AI agents and you need some books let us know We might we might be able to hook you up No promises but we we can talk about it All right So yeah let's just I'm gonna open the book and maybe we should

just I'd like to talk about a couple things First let's talk a little bit about some of the things in the book You know what what were we thinking when you know you wrote it and let's also maybe talk about what's next because one of the things I always hear you know when we give people this book is "Isn't this out of date the minute you print it?" And so let's talk about that and what

we're doing to try to like help uh help prevent it being out of date the moment it goes to print Well so the first thing you do when you're an engineer is you set up your your your data pipeline right and so we we discovered that uh you can like Amazon has a printing service and you can just upload new copies of the book whenever So I'm

actually working on V2 over the weekend and we should have it live you know pretty pretty soon on Amazon Um the the the origin story for everybody watching like Shane and I were were sitting in this very apartment about like three months ago and we were just kind of like we we'd been bringing folks to the

apartment to do like these whiteboarding sessions where people would kind of like go and we we would like here's our here's our whiteboard Um you know here's our whiteboard So on that whiteboard um on on that whiteboard we were uh you know we would just kind of like have people whiteboard out their architecture of their AI application or agent or

assistant you know where's the data coming from like what are the decision points you need to do and and usually over the course of a 45 minute uh session we would kind of have like two or three key unlocks where like oh well what if you instead of making one LM call to you know maybe what if you broke

to like diagnose 12 symptoms what if you broke that up into 12 separate LLM calls or like you know something like that Um and and over the course of having these like 20 or 30 different conversations it just kind of felt like a lot of what we were doing So some of the stuff was like very specific to the particular use case

Um but some some of that we were having a lot of the conversations over and over again And it's a normal thing when you're an engineer writing docs right it's like if you have the same conversation over and over again yeah maybe you should be putting that in the documentation Um but then we're like well what's the documentation for for

building an agent like okay maybe we should write the documentation for building your agent right like and so so we kind of like think about this as like the documents and like you know docu if you think about like what is you know you can you can kind of break down content into like the how-to material like here's the actual code you know that you would use to do a thing but

there's also like the con there's also the conceptual like what is the concept here that this code is like the reference implementation of and so we spent a lot of time just kind of like uh like thinking about um like Hey like what are all and like luckily we' actually thought about this for master too we have like you know what are the primitives of AI engineering agents and tools and memory and rag and evals um

and tracing and you know etc etc right um so like and then like prompting and and like what about models and providers right like you know um and and so like we we we already have like we already have a very like clear sense of like what all the primaries are and then we kind of like codified that into sort of

a table of contents and then like kind of generated then sort of like we're able to like I just kind of like locked myself in a room for like a couple weekends and then like wrote uh sort of like wrote the book basically Yeah And it's funny you mentioned you during YCS having all these people over to do whiteboarding sessions So I'm going to

share uh we actually you know we had so many people that we uh oh yes we built an agent that took a whiteboard picture and turned it into an Excal And unfortunately we built it towards the end of all these whiteboarding sessions but as you can see here like we have a bunch of pictures from random startups

coming over and just you know whiteboarding with us at that at the Palace of the Dog in the Dog Patch in San Francisco So yeah there there you go Um um so I do Yeah So I do have some questions on uh specifically on the book and I think it kind of ties into you're doing a talk at uh what's the conference you're doing a talk at here pretty soon

ai.engineer uh the world's fair AI engineering conference in SF that our old friend Swix is running Um 3,000 folks are going it like a lot of you know Greg Brockman is speaking a lot of like really cool folks are speaking I'm also speaking Um the topic is agents and um and workflows Why not both um OpenAI

had a blog post that was kind of a little bit uh you know not very uh friendly towards if you there's like a speakers I think you're for whatever reason that's like the mobile version If you click speakers then and scroll down for a while then you get to me Uh uh they had a um so I was actually just talking and Dex was like oh you're you're gonna be there Dex you're going

to be there and you're he's going to be there too Um I have a lot of friends speaking at this conference as well Um uh the the talk is called agents or workflows why not both cuz one of the you know there's been a lot of controversy controversy around like um you know is an agent just a workflow graph that you chain things together or is an agent more than that um and some

people are weighing much more on one side Uh some people are weighing much more on the other side Um and I think the talk is just kind of like you need both of these things Like this is a dumb argument Yeah And I I agree with you I think you know it's it made sense for people to lean and I think it's a it

changes over time right like it's it's a sliding scale where maybe you know six months ago you almost had to use predominantly workflows to get anything done Now it's you still need to use a lot of workflows but maybe you can turn more control over to an LLM to make decisions and and route things And then maybe over time you can turn you you're

less reliant on workflows and you can even turn more of the decision-making process over to just a really well-crafted system prompt Yes I 100% agree with this And I also think that you know if you're already spending all this time writing this really detailed system prompt of do this then do this it's like why even leave it up to chance if if you can get that specific if you always do these things

under these conditions why not just turn that you know especially if if you can use AI to turn that into a you know a more discreet workflow Yeah we have those AI tools that they that can help you write the code for it Yeah And I think like what what people often don't realize is that like there's different types of agenda capabilities and you need them like differently Like I was

just talking before this call to someone who was like "Hey like do I need a um a planning loop?" Um and so a planning loop is where you have like an agent running in a while loop and the agent can create tasks and those subtasks and so on And I was like well we see some people using it The people who are using

it are usually building coding agents If you are not building a coding agent you are unlikely to need that sort of full flexibility um in production that full flexibility kind of like agent planning loop um or you're sort of like less likely to do that So like it's it's kind of like but like this is the kind of nuance that

like maybe we see because we just talked to a bunch of people building different kinds of agents that you might not realize when you're getting started with it And so like this is the kind of color and detail that again like we we kind of put in in the book Yeah I think it's I think it's really useful and I think it's one of the conversations that I have the most is like what how much control to give the

agent versus when to use workflows And how do you think about I know I know I like I have suggestions I give to a lot of our customers but how do you think about if someone's just getting started should they start building a workflow first should they start with an agent should they start with an agent network i mean or a network of agents How do you how do you think about that i think it's

useful to sort of like to play around with all all three things right like before going really deep on on each one If you think if you're not sure where to land like generally the way that we that like I tend to advise folks is like if you're sure you need like you know really fine grain control use a workflow If you want like a lot of power like go agent or multi- aent right it's just a

sliding scale of where you want to be on this scale Um yeah I often tell people start in like if you don't know just start with a single agent and a and a good prompt and work on the prompt and then you're going to probably quickly realize that the accuracy isn't good enough for certain things and then you pull off those certain things into workflows or specific tools and you can

kind of like slowly kind of slide the scale or if you or if it's you know you need it to be even more flexible maybe you start to imple you know add more agents to the mix It's almost like start in the middle and then like slide you know slide a little bit one way or the other depending on what the tasks you're actually trying to accomplish One of the things that people say they've

appreciated most and they were like surprised most in in uh the book was we we gave like about I don't know 20 or 30% of the bold new uh system prompt and it's just this incredibly detailed um you probably can't read this but like you can just kind of see by how small the text is and how how long how long it is like this very incredibly detailed

system prompt Um and and I think like that's surprising to folks because um you know people think about oh yeah I wrote a couple paragraphs and like in production you might write a couple pages Um in fact I was just on a call with a someone I can't we can't show publicly but they're a a prominent um let's say they're a prominent startup shipping a um an agent um shipping an

agent and they they sort of like at one point in the call they like screen shared part of their their system prompt and and it was like again it was like a page or two p you it's a it's sort of like it's a it's a um the detail and depth is is detail and depth in the system prompt is is important right yeah Yeah Absolutely I I think that you

know ju just like if that the system prompt is kind of you have to spend like an excruciatingly large amount of time on that one specific thing kind of like if you were right building a whole system but in this case a lot of the system is is kind of codified into one prompt which could be very large right which is also I I I think one of the

challenges too is there are of course prompting guides there are some techniques you can do to ensure that you know you get the right prompt alignment and all those things but it is a little bit like the wild west you like people are experimenting with things and finding out that certain things work and certain in sharing those tips and it's a little less uh because it's not

deterministic it's less defined and and less clear how to you how to structure these problems so the best way is just look at other examples of how uh how people are doing it then actually having success with yeah 100% agree and I think like one of the things that we you know there's a lot of intricacies that might be specific to your situation Um but there's a lot of things that are general

So I think I just encourage folks to you know sort of like hey ping us in Discord right or like ask us like hey I'm doing this particular thing And you know it's it's like humans are the same as agents The more context that you give like the better the responses right and provide like provide a lot of context about what

you're building and ask like "Hey like do do you guys think this is the right architecture?" Right ask us on Discord ask us on Twitter like you know just talk about like we just you know just talk about what you're building public with us and public with others that like have AI engineering experience Talk about your architectures

like and and like you we're all like I think what one of the other big lessons um that sort of like crystallized from the book and from YC in my mind was that like a lot of us are just doing this together we're you know there are a lot of very talented engineers who are um becoming you know AI engineers and and

it's sort of like it's just a new domain the same way as DevOps used to be a new domain the same as data engineering used to be a new domain There's some weird things and some intricacies but at the end of the day it's just a new domain And there are unique things about the domain that you're kind of like learning and picking up along the way Um that

like you learn the fastest when you're learning in a community of of practice Um you know or or a community of like a community where you're all kind of like building together And if you're sort of like out isolated on on an island by yourself like not reading blogs not reading sort of like not using tools other people use you're just going to learn slower than

everybody else Because the only reason that people who are learning fast are learning fast is because we're learning together Yeah absolutely Which which is why if you are looking for people to learn learn alongside go to master.ai and scroll down to the bottom click the Discord link and just come into the

Discord and tell us what you're building You know we'll we'll try to help you out We're all learning this together We're building things on shaky foundations right now right the whole space is moving really quickly So uh we're all trying to keep up and all trying to build really you know really interesting agents and agentic applications on top of models that are changing and tools

that are changing all the time specs that are changing which I guess uh follow-up question now What are some of the underappreciated parts of the book do you have any topics in here that you think people are either not spending enough time on not uh haven't either got there yet or maybe they're they're not focused on enough i think

underappreciated parts right now is rag Rag is not that um contrary to like expectations I was just you know this is just a dumb example but I was searching on Amazon my Amazon order history yesterday uh to try to reorder my trash bags and I searched garbage bag and nothing came up I'm like I swear I've

ordered this And I deleted garbage bag and I typed in trash bag and it showed up in I like reordered my come on this is Amazon like it's semantic search on my order history and like Amazon hasn't got this right like you know like rag is not dead because like this is a dead obvious use case and the biggest company in the world with the most resources has not yet you know shipped it right rag is

not dead um so I think like the the the chapter in rag is is like useful to just like think about like you know if you have search in your application how could I have semantic search what would that look like if I had semantic search in the application that I'm working on yeah I I agree with that I mean come on

Amazon let's get someone from Amazon on the stream let's let's talk about getting uh semantic search added to the to the order you know it's you gota you'll have to pay a few tokens maybe you know but I I mean it's also like it's also just funny because like they have Amazon bedrock right like it's not Yeah they don't have the infrastructure to do this stuff Anyway come on That was

But that's that's one of many examples Yeah I would agree with you Rag is not dead But on the flip side I'm finding more and more people are able to push off moving to Rag for longer Yeah Yeah because because there's also something real really interesting too where it's moving towards like and we'll have some of this in the the new edition of the

book like you know tool augmented generation where you sort of give an agent you you know access to tools to fetch data um and and so like you could think of you have like a medium it's still rag right I mean that's still it's still rag but like technically it's rag but I do think it's like different

enough that like we kind of coined this term like tag like tool augmented generation We also called it MCP augmented generation which to be fair was like really just us feeding the Twitter trolls and like we should really call it like tag instead of mag Um but like uh yeah um to like yeah our investor Elena Goyle like published a

blog post on like her personal website MCP server like I would go take a look at that if you're kind of curious when you have like a medium-sized corpus of data like a a website or maybe just you know something then like you just feed it into Gemini and you don't get any good like analysis on it and you want to like give an agent some tools to like

analyze it in a meaningful way tag is kind of really nice Um yeah absolutely I think that as you know as you can just turn over some of the control to the agent to do its own research whether that's using a vector database whether that's using a traditional database or some other kind of you know search tool right it's providing a similar you know it's trying

to do the same things It's just a different mechanism to get the data into the context I I will say you know so I did was talking to someone at a meetup last week and you know they had compared a rag solution compared to just using uh Gemini's long context and they you know I would say they got less latency just sending a whole bunch of um just a whole

bunch of data into the context window and so they were able to like decrease latency and increase accuracy by just using more of the context So I think it does depend There are there are cases to be made for for both but I think it's no longer I think rag was just a thing everyone grabbed right away as soon as you need some kind of outside knowledge And now it's well it depends on the situation and you honestly it's a little

bit more nuanced and sophisticated these days Yeah And you and you have to do you have to probably test you have to test the assumptions right try probably have to try both approaches and depending on what your goals are you'll you'll see what which one which one will fit better uh depending on the amount of data you're dealing with and and what types of data Yeah Oh well I I gota I got to

hop to a one o'clock unfortunately So you may want to um phase me out and uh bring on the smitheries Yeah Yeah We we'll have another guest here joining pretty soon And yeah thanks for joining Sam and yeah we will talk to you later Okay Good to see you everybody here in the street All right Sam thanks for

joining and uh I think we will have Henry from Smithery joining here soon So we will get him queued up and joined in For those of you that are just joining we have been talking about the Maestro.build hackathon which is going on now It's not too late If you It's going on all week so if you wanted to build some kind of agent everyone that

uh submits something will get a copy of Sam who was just on his principles of building AI agents book We also talked with Steve from mcp.run We built an agent that can do research can go out and find AI startups that have gotten funding It built an Excel spreadsheet it built a chart and we did that in you

know about 25 minutes or so We built that whole agent and got it running and we'll try to get the we'll we'll post in the comments if I can't get it from him before the end of the stream We'll post in the YouTube the Twitter and LinkedIn comments so you can have a link to that code if you want to pull it down play around with it on your

own And we do have some uh comments that we can kind of follow up on while we are waiting for our next special guest So we have this comment from Adabio I believe I'm pronouncing that I'm not great with pronunciations so hopefully I got your name right Uh can you throw some more light on the MRA client SDK so we do have a MRA client SDK and it's essentially allows you to

you think of when you're building a MRA agent or MRA workflows you essentially get a REST API server you're deploying these things in you can deploy them in a completely separate environment So you have almost like an AI backend If you do go down that approach then you we have a client SDK which you can add to your front end so you can more easily

communicate uh with your backend agents your backend workflows you can trigger a workflow watch its results show something in your UI uh all through using the client SDK Now there are other approaches A lot of people also you know they have maybe an X.js site and they want to bundle Maestra within that They don't want to use it as a separate API

They want to actually just bundle it in the code You can do that as well And in that case you don't necessarily need the client You can just interact with the code directly Just call you know just load your agent call agent.stream workflowgenerate Um and you can trigger all those things without having the client So the client is really if you're

kind of going with a kind of separate approach where you want to separate your AI uh infrastructure or your AI backend from your front-end client And this is uh we do have a lot of customers that are building you know desktop applications uh that are potentially building mobile applications or maybe have a a website and a mobile app that are communicating

with the same agent And so this kind of client helps especially if it's you know a JavaScript or TypeScript type project where you can just use the client and interact with those things And let's see what else we got for questions here So a few comments here This one's as the system prompt gets long and detailed I've seen that AI agents don't really

follow the path So as of now we're sticking to good workflows decided by agents when to pick which workflow Yes I've seen that same kind of thing where you eventually essentially you have like a router agent and you give it a set of tools which maybe are specifically like well-defined workflows and it can do a good job deciding which

workflow to run but maybe it you can't give it all the tools in the tool belt right you have to have more defined discrete steps in those workflows and so I've seen that a lot with you know some of the projects that I've been working on but also a lot of the customers uh that I talk to as well and with that we do have uh Henry

from Smithery joining us So let me get him pulled in For those of you just joining we've been talking with a couple guests If you know of someone else that we should be talking to on this live stream someone you'd want to hear from uh let's let's make it happen Send me some information Make a make a referral You can find us online We we are

around All right Henry are you there hey Yes Uh can you hear me i can hear you Okay that's great Yeah the video is a little bit broken up but it looks like it's starting to come in now So hopefully everyone can The audio is coming through great though Awesome All right Henry can you maybe a quick introduction would be would be a good thing And then we'll talk about some of the things we can maybe demo and then some Q&A I think would be great to

have as well Cool Well I do have a slide deck that you can bring into uh you know view if you want me to kind of show that Yeah share if you want to share it Yeah I Okay I put it as share Yeah there's a there should be a share and you can just share your tab or whatever and we should be able to pull it up and everyone can see what we're looking at I I already shared it I think you have to as a host

you have to let it in Oh yeah Well we're still new around here There we go All right Um okay Hey everyone Uh uh hold on All right Went to the back Okay So um my name is Henry I'm the founder of Smithery Uh I had a previous startup I exit from that Um and you know today I want to talk a little bit about sort of the origin story of Smithery uh briefly and then show a little demo of like um

what you can do on Smidy uh as a developer So this all started you know last year u towards the end of last year when I was playing around the arc agi challenge um I was trying to solve this uh kind of AI benchmark puzzle um using uh reasoning models um and um you know as I was trying to solve this uh challenge uh you know one one interesting thing about this arc ai challenge was that um it's very easy for

humans to solve uh if you look at this puzzle it's pretty easy to guess what the third uh sort of prediction should be but it's very hard for modern uh LLMs to solve Um so uh towards the end of last year when 03 kind of released their benchmark um they completely destroyed this puzzle Um and I was like okay I

guess you know AGI's you know sort of achieved we have reached like you know human level performance on math Um you know we're kind of done right like we can just pack our bags and all go home right Um and so when January came in 2025 I was like looking around and asking the question of like well if AGI

is here where are all the autonomous agents um and so what we really have is like uh this Claude's paradox right uh we have like models are getting smarter on these benchmarks but they are not connected with the rest of the world Uh they're kind of disembodied like this picture uh I'm showing here Um and what

we really need is that we need to provide models with the right inputs and outputs Uh and this is essentially what uh MCP is like MCP was established in November last year I started tinkering around when MCP uh you know was just released It was a small but vibrant developer community back then Now it's uh it's a lot more mainstream but you

know it's still very early The protocol is still being developed And MCP kind of promises to be sort of the USB uh C for your AI agents allowing it to get inputs from the outside world and outputs uh applied to the outside world Um the problem is that the MCP doesn't really solve all the problems here Um it's kind of a fragmented ecosystem as some of you

who have dabbled in MC might have known Uh it's there are a lot of MCPS out there You kind of had to go on these lists uh back in around December to kind of find MCPS uh installation is pretty hard and also like it it could have like a virus somewhere Um so you know that's where smithering comes in You know we're trying to solve this problem of

defragmenting the ecosystem of MCPS by providing AI gateway for agents to access all the tools and context uh in the real world Um so our website well this is an old old screenshot but I'll show you in a bit Uh our website gives you kind of a curated list of MCPs um as an end user to use but also as a developer which I'll show in a bit Um

and yeah we have we're processing about 50,000 tool calls a day now Um yeah so let me jump over to uh change my screen into a uh to make it more of a demo Um can you allow that okay it's working So um you know this is our landing page uh it's kind of made for both users of MCPS and developers Um so let me just show you uh what it looks like if you want to use an MCP in your uh agent uh application So

let's say you go to an MCP um uh let's say we want to add search into our kind of um AI agent uh through MRA of course Um what you can do is you can explore the catalog MC MCPS on Smittery You can play around at tools For example I'm looking at EXA X is a search engine You can play around with the tools So this is these uh these tools are like what your AI agent would would see in this

tool list So for example I can search you know for any query here and the XA API will uh process this through the MCP Um it can then uh return the MCP and um give you this result Uh this result is basically the string that your AI agent would see Um so to integrate this just click on the API tab Uh we have some SDKs that you can use Um we use the standard stream bowl uh HTTP transport

So uh it should work uh pretty smoothly with MRA as well Um and if you're on Python we have a tab for that but we recommend uh you know TypeScript I mean we're not we're not anti-Python but we're definitely protypes script you know Yeah I'm also protypes script Yeah Can you can you zoom can you give one

more zoom like one click of zoom on that because it is just so everyone can can see it that that's a little that's a little more helpful So uh it's kind of cool though So you're saying you have this interactive playground so you can kind of test the tools before you hand them to and you know whether it's like cloud desktop or cursor or obviously if you're building an agent with MRA you

can you can test it here make sure you're getting the results that you want and then pass those or turn those on and pass those in Yep Um yeah I guess before I go on anybody have any questions about how to use MCPS on uh that are deployed on Smither yeah if you do if you are tuning in live you know we are live So whether you're on LinkedIn whether you're on YouTube whether you're on X just drop a comment We'll see it We'll

answer some questions along the way And if you have questions on you know how to use MCPS with Smithery If you have general MCP questions bring those up as well Happy to field those I I'll see any questions yet but I imagine you know in in a few seconds we'll probably get a couple So let's we can probably continue on and I'll pull up any that we get

along the way All right Well feel free to interrupt me when you uh see a question Um so um I guess moving on uh if you're lazy and you don't want to pick what tool to use um I we've built this uh uh tool inhouse called toolbox And toolbox is just an MCP that actually calls all the other MCPS on smittery And uh your agent can uh use this function

to search the entire registry to figure out which tool to use And you can your agent could use uh this to dynamically invoke any tool on spittery Um and similarly we have this uh kind of code snippet to integrate into your agent Uh so that's kind of how spidry will work if you are kind of consuming MCPS as

part of uh part of building your agent Uh you can also deploy MCPS as a developer So let's say as part of this hackathon or maybe in the future you want to uh submit an MCP into the registry uh we have sort of a versell uh-ish experience where you can you know find a repository uh for example I will connect to some you know a reference

sort of server that I developed on uh we've developed on github um let's call this you know github 3 and uh we can click create and this will what what this does is that it would sync into your github repo it will start running a cicd and deploy and host your uh servers automatically so this takes about a minute and once it's done you have your

own server page um that you can kind of give to other people and allow other people to use your MCP So um in summary we have like two ways to use MC uh two ways you can use Smittery as a consumer of MCP uh or as sort of a developer trying to help other people by creating your own MCPS Um yeah any uh any questions you caught

Shane uh so so I I have some questions So one can we is it possible for us to show this hooked up to I don't know cursor or claw desktop and just show how easy it is and I know a lot of people are building agents right so if you're building an agent you're going to pull in MCP but also I think there are you

know maybe some people who aren't as familiar with MCP don't understand that a lot of the tools you might already be using there are useful potentially MCPs that you could add into those tools to help uh help them along And one example of uh an MCP would be we at Maestra have a MCP doc server and what that is useful

for is if you are building a MRA agent in cursor or windsurf or VS code that you know any any tool that supports MCP you can just add that in and then if you ask qu specific questions about MRA it'll actually look up the information on the docs and it stays up to date So the benefit there is you know as we know

it training on these models happens at a point in time right and so depending on when the training happened you know it's only going to know a certain amount of things about MRA or any other tool and so by having a docs MCP server you're always getting up-to-date docs and it can do a much better job helping you write code

Yeah Um so so yeah you wanted sort of a kind of a in the client sort of demo right So yeah it would be great to show just you know whether whether this is you know this could be cursor or obviously we're looking at what cloud desktop here but actually work Yeah So uh I guess one example I can show is just now I talk about toolbox right which is like an MCP that could search other MCPS So I could

ask questions like you know um you know what's happening uh today in SF um and uh it's kind of a vague question so we'll see if it works Um usually usually it's useful to have a a prompt that like teaches um your agent how to use um MCPS Um so uh let's see this works So with this uh prompt um ideally what it should be doing is um it will kind of check uh for a tool that can handle this request

Um okay we need to allow allow this Um so in cursor you will have a similar kind of interface where it will search uh for servers if you use toolbox or any other MCP And let's see All right So um it did a first pass search on submittery Uh it was able to find a couple of MCPS Um let's see Oh wait Hit

the maximum like Okay I think the contact window is busted um for this particular uh query because big guy did not pay for claw pro but yeah Yeah I mean yeah no no worries I think but the idea here is you were able to add this toolbox this to cloud desktop ask a question it was able to then execute that tool depending on the context of

your question and in this case go out and do a search but it could be you know could be many okay I have a I have a recorded video demo which kind of worked in the past um so in this example uh same example here uh let me skip forward a bit um so this is a different client that uh we kind of uh prototyped during

a hackathon Um so what happens is that uh it would um be able to find sort of MCP that's relevant for this particular use case Uh so this is kind of like an agent chatbased agent uh is able to use the tools and um you can also uh post this uh you know you can also connect to Slack So this is a different MCP like Slack MCP So you can imagine agents

being built to be uh kind of connected to different services different integrations in order to actually uh solve real problems Um so that's what we're hoping to see um you know kind of people built on top of uh Smittery Yeah Do you have any you know so there's a few uh general questions about MCP but

I I'm curious what are some of the you know you you obviously talked to a lot of people using MCP and using Smithery to build uh you know interactive AI applications or you know agentic applications What are some of the use cases you've seeing you've been seeing that are exciting to you so I think the most um uh so I think you know obviously there there are the integrations which are kind of you can

think of them as like foundational building blocks uh without integrations you don't have access to a lot of apps but uh perhaps the more interesting uh MCPS are the more AI native stuff like the stuff that did not exist uh as a traditional API right before MCPS existed uh one example u is sort of this

u this like clear thought server um what this MCP does is that actually it doesn't provide you with any integration but it actually provides you with a bunch of like thinking frameworks So when when your when cla or when your model goes and use us use these tools it kind of constrains the the kind of way the model thinks by by for example in sequential thinking it kind of forces

the model to do a dynamic and reflective problem solving So you can think of this like as a sort of like prompt++ because the prompt is actually also a program that kind of guides the way the model solves something So this is kind of useful if you have a predefined way of solving a task right that you want the

model to use Uh another obviously another example is like things like more AI native like browser based right you you're able to now get AI to control the browser Um so yeah those are kind of two examples I would like to highlight that are um pretty uh cool in this uh more AI native Yeah speak Yeah definitely We

we've seen a lot of interest around agents that can interact with browsers right so you have browser base you know shout out to Paul We need to get you on this live stream at some point Um you know browser use we we know the team around there as well But some really cool interactive ways where you know your agent isn't constrained to just

your environment they can actually use a browser and you know whether it's buy something from you on Amazon which is with the demo that I I tried to build and it worked pretty well or you know whatever else you know humans use browsers for right agents can interact with browsers so I see a lot of like interesting MCPs popping up around like

trying to use browsers with agents uh awesome yeah anything else that uh that you wanted to chat about today Henry I'll look see if we got any other new questions that that have come Yeah I guess to just wrap up uh this uh little demo and talk Um you know obviously the goal here is to uh help all of you agent builders unlock today's agents to be able to do more stuff but I think another interesting thing um

perhaps more more of like Spitter's vision is like we also want to help help bootstrap tomorrow's agents um you know there there's this um recently written um essay by uh Richard Sudden uh and um and a co-author uh on how like we're entering this era of experience and it's a very well-ritten article because like it kind of talks about how all the LMS today they've been kind of trained on

data on the web which are sort of like human outcomes right uh but what we want to be going towards is a world where agents are learning from their own experiences um and sort of kind of uh getting agents to like really uh do stuff in the real world and collecting that data and kind of learning from that data I think

that's perhaps the first step necessary in order to bootstrap tomorrow's agents So you know um this is like um the era of human data where you know agents are kind of learning from static information but you know we want to get to a point where we have an era of experience Um so yeah and is that the name of the is that the name of the paper era of experience

yeah the name of the paper is era of experience I recommend reading it Yeah Yeah I see it Uh yeah I just did a Google search Welcome to the era of experience by David Silver and Richard Sutton Yeah All right I I have some light reading to do now Awesome Uh well Henry appreciate you taking the time to come on Uh tell us a little bit about Smithery a little bit about or you know show some examples of

how Smithery works you know there's a lot of excitement around MCP and you know we're we're trying to you know help help others learn alongside of us and so appreciate you kind of teaching us a little bit today Awesome Yeah Thanks for having me Yeah All right Henry We'll see you later All right everyone We're about to

wrap up For those of you that have just joined you know we've been talking about the master.build hackathon It's still going on If you would like to build your first AI agent with MRA if you've been thinking about it but you haven't committed the time to do it yet but are considering it this week is a good time It goes on till Friday So there's still

time for you to build something Go to master.build to learn more about that to get signed up Go to our Discord where there's a Monster Build uh channel We're chatting about all the cool things that are getting built If you watch the live stream yesterday you'll see a really cool example of what someone is building And we'll also have some special guests

on tomorrow as well uh showing what they're building in the hackathon this week So more more to come there We also talked to Steve from MCP Run We built a little research agent with MRA and MCP Run We'll share the code to that in whether you're on YouTube or LinkedIn or Twitter X whatever we call it uh we'll

post the link to that code here shortly So follow up and we will get access to you for that So if you want to play around with it and build upon the simple uh structure that we built which was just go out do some web research finding the AI startups that have received funding recently add it to a Excel

spreadsheet and build a chart around how much funding they've actually raised We talked to my co-founder at MRA Sam about the AI agents book Principles of Building AI Agents If you want to grab a digital copy you can do that Just go to let me just put it here master.ai/book Go ahead and go there You can get a digital copy If you want a physical copy it is on Amazon Or if you

submit anything for the hackathon you'll get the book So we'll send it to you And last but not least we also talked to Henry from Smithery which again more information on how to connect different tools to agents That is something that everyone's trying to figure out because once your agents the once you can give

LLM's tools it can unlock a lot of different things So that's it for today Hope to see you all tomorrow Uh hopefully we'll hear back from Obby Uh he should by tomorrow he'll have landed in Japan So we'll try to get an update from from Abby who's not going to be here for a little bit but hopefully he'll be doing some live streaming of his own while he's in Japan Uh talking

to a lot of different uh a lot of different developers that are using MRA in Japan But that's it for today Uh if you do have any other uh guests that we should bring on any ideas let me know Reach out You can find me on X go there follow me uh reach out My DMs are open And we will chat with you next time See you later