A "Course" as an MCP Server, Lauren from Ampersand, and AI News

I'm back and yeah we have a great uh great show for you today we're going to be talking some AI news like we typically do we're going to talk about the new MRA course we released last week but not just you know we're not just promoting it we're going to talk through how we built it we're going to look through the

code and see how it all works and show an example of it actually in action and then we're going to talk with Amperand and talk about the Maestra integration that they just released and talk about their MCP server and we will learn a whole bunch of uh learn how to connect a whole bunch of tools really quickly to our AI agents that we're building

as always you can make sure you're following all the MRA content you can also follow me on X SM Thomas 3 and this is a live show so if you have questions please drop a comment wherever you're watching this whether it's YouTube LinkedIn X and I will try to answer questions as we go but with that let's dive into some of

the news for today and uh some of this is probably review from some of the content that we talked about last week but I did just want to highlight uh because I think it's pretty cool so the first thing let me share my screen here you've probably heard about this i'm just going to talk about it briefly but if you haven't if you uh have not

watched the stream recently or you have not been paying attention to all the you know the things posted about it last week on X and other places Cursor released their background agents which is really come off the backs of a lot of these other tools like codecs and like jewels and even you know GitHubs kind of agents the idea is you can run these

things in the background so you can basically spawn an async agent that will edit and run your code in a remote environment so rather than actually running it you know local on your machine it's running it remotely similar to you know how we've kind of demoed on the past about codecs and how Codeex works but the cool thing about it is you

can kind of uh spin up these agents let them run review their work and then decide if you want to you know go ahead and merge it as is make changes whatever uh but basically you just use command E it opens up the background agent control panel and then uh you can go ahead and submit a prompt and just let the thing run so it's kind of built in right into

your IDE um one of the questions I guess that I have is I do think that these background agents there it's going to happen where there's going to be some winners right it currently doesn't work for all tasks but last week one of the cool things that I did is I was using codecs quite a bit and I was able to just from my

mobile device I spun up like half a dozen or so tasks that I had on my mind that I knew I wanted to do but I wasn't going to have time to get to because I was you know I guess quote unquote on vacation if you can call uh me still uh spinning up tasks for Codeex really on vacation but I was uh just spinning up you know half a dozen tasks and then I

would come in once the P once the it was ready and I'd have it create a PR for me and then I just review the PR and I merged a few a few of them I uh just left because I wanted to come back to but it was a really interesting way to think about how you're just in some ways delegating work at least I will say the

more trivial tasks the things that you would know how to do well that you can somewhat describe I I feel like it does a pretty good job on those types of tasks and I imagine that uh cursor background agents would be uh very similar if you have certain refactors or certain things that you want to do and you know it's just kind of tedious but

you're pretty sure you know how to do it i think you can provide enough detail fire it off and then just review the code rather than you having to write the code yourself so curious if anyone's used cursor background agents and could compare it to how they how their experience was with codecs or any of the other tools but that was that's our

first uh first story for the day next up there has been uh a little bit of I'll call it some uh drama on X around this research report or this paper that Apple has put out called the illusion of thinking and it's it's really talking about uh reasoning models and some of the limitations of reasoning models so here it is on the machine learning.apple.com apple.com site but it

also we can actually look at the full report we're not going to go through the whole thing of course but it's here you know it's a pretty you know pretty solid 30 30 page report or 30-page paper so it does I think what you know my takeaways so from I haven't read the whole thing but I've read pieces of it and I've kind of got some summaries guess there's really only

you know if you get through there's only 10 pages of actual content um it really kind of throws uh throws a little shade at the idea of you know quote unquote reasoning and say how it actually does its reasoning and basically it talks about how it does break down you know when it does its thinking i would say this doesn't

surprise me i do give Apple credit for being the ones to kind of like you know throw you know throw cold water on the fire a little bit i do think that they're you know I've been saying for a while I think others you know have been saying that at least questioning are is the current generation of how LLMs work

really going to get you know get to this deeper level of reasoning that might be needed to achieve something like AGI and you know I'm not a machine learning researcher i'm not I I don't know the answer but I do think that it you know if you've used these uh reasoning models you kind of understand a little bit of how it works and you've probably seen that it does break down for some uh

deeper tasks some really complex tasks i don't think any of us are probably surprised by that right but I would say you know I've seen a lot of people saying that um this is you know kind of proof that reasoning models aren't going to work and and I don't think that's the case so I think you know in some ways Apple is maybe you know putting out some negative

negative things around uh some of these LMS because honestly I think Apple's kind of struggling with how to deal with AI and you know if you if you've used Siri you you'll understand i think they're they're way behind in my opinion and uh even though the large reasoning models are not maybe as great for the

really complex tasks as what you know we would all hope they'd be you know we want to see them continue to improve i do think that if you just imagine even if they I always like to just put myself in the shoes of if none of the models got better than they are today i still think there's a lot of amazing things you can do with these

models so they will of course continue to get better how much better some people think it's going to continue to accelerate others people other people think maybe there's a natural plateau time will tell but uh I'm not surprised that you know they're questioning how reasoning works in these reasoning models and that they're saying it's not

really doing true reasoning i think that's kind of understandable it's just layering in next token completion in a lot of cases but still interesting next up let's talk a little bit about uh something else that was kind of re released or at least talked about i think I heard about it last week but maybe it's getting announced more broadly now is that OpenAI has updated

their voice mode to be to have a little bit more natural sounding speech so over the weekend OpenAI rolled out an update to advanced voice the feature that lets users have dialogue with chat GBT out loud so if you've used advanced voice mode you know it's pretty good uh but it's not always perfect they you know they do have a realistic apparently more realistic cadence i've also heard you know I've

seen I think it was Swix was trying to sing with real uh with advanced voice and it was kind of bad but it worked you know it honestly did try to sing a song so you can uh you can see that they're continuing to lean into making the voice better i think this is a very natural way that people are going to want to talk to these assistants rather than just type messages that way you can be

hands-free you can you know it can be what Siri uh maybe was supposed to be and so encourage you all to try that out curious on how if anyone's seen any improvements i haven't used it enough over the last weekend to know but I do use advanced voice quite a bit and so I'll be interested to see how how much it has actually improved how it sounds and all of that

all right so for the last I guess piece of news that I wanted to discuss this one's more I would say a discussion so we'll spend a little time kind of talking through this and it's really around MCP and we're going to actually spend quite a bit of time talking about MCP today but this was a post from June 6 written by Frank Feele

and it is MCP versus API so just the quick summary every week a new thread emerges on Reddit asking about the difference between MCP and API and honestly I've seen these two like why do we need MCP why can't we just use API and isn't the open API spec good enough um so he lays out quite a few uh reasons

why MCP is a good thing for you know specifically for using within you know AI tools and LLMs i think one of the things you know is that the API you can with APIs you can have a whole bunch of there's not necessarily set standards right you can have different types of APIs you can encode them different ways um so you can see there's different ways you might

encode it with URL pass request headers query parameters request bodies and you can document it with open a API or swagger documents but it does uh it it isn't necessarily considered consistent right not all APIs follow very consistent standards they try to document those standards but you do know that there's a whole bunch of different ways that companies will build APIs so where the Frank's argument is

that MCP is more of a wire protocol that helps enforce consistency so there's two different types of transports you can run it locally with stdio standard input output uh or you can use streamable HTTP it has discovery more natural discovery so it has this idea of tools and resources so you can get the capabilities of that MCP

server at runtime it has a set of primitives tools resources and prompts and you know there's more there's some you know they've recently introduced more around O as well at least in the draft state so there's a whole bunch that the spec is moving for sure but you can of course check that out so the question is why not just use open API so the open API describes while the

MCP prescribes i think this is very a very good way to put it mcp is way more prescriptive so it's not just about documenting the way you do it it's saying no if you want it you MCP if you want to use MCP you have to do it this way um and yeah maybe open API could add some of these things but then you have to worry about adoption sometimes it's better to start fresh so then we'll talk

about the five fundamental differences so there's runtime discovery versus static specs so with an API you're you ship new client code when the endpoints change but with MCP your agents can query the capabilities dynamically and then adapt automatically i would say that that's this is the dream it's not always true

in my experience like a lot of times you need to have your a depending on how you write your agents and how you build your system prompts you do kind of need to know the tools ahead of time for it to really uh react well maybe that'll improve you know as we kind of mentioned earlier assuming models continue to improve maybe we don't need to do that but I'd

say this is at least the dream we we I can see us maybe getting there at some point where with with API with just standard APIs it's not going to happen uh deterministic execution verse LLM generated calls so an LLM writes the HTTP request so I guess saying that you know APIs and MCPs you know handle this slightly differently so in the MCP scenario LM picks a tool and

then that wrapped code executes a little bit more deterministically because it is just kind of a set tool rather than just exposing your LLM and giving it access to querying the AP the API i think you can be a little bit more structured like not every API endpoint maps to an MCP tool i think sometimes you'd want to combine multiple API endpoints into a

single tool for instance if you're thinking about the whole like a an a checkout process there might be multiple tool calls like add to cart insert payment information there might be ways where you wrap multiple API calls into a a single MCP tool call and you can make that more deterministic birectional communication is a first class feature of MCP i think

that one's kind of explanatory apis kind of have this depending on what what you're using but it's not necessarily always standard uh so REST APIs fragment human task across endpoints yeah and this is kind of what I was just saying before so I guess I was maybe jumping the gun he already covers this but yes like MCP tools are more closely mapped to workflows which

makes sense like you would want to give your an agent an entire workflow not just a whole bunch of different uh things that they have to piece together themselves and then MCP's local first by design i mean originally it was almost all MCP servers were stdio with didn't need any network i think over time more and more will use streamable HTTP and uh

so maybe this becomes less of uh you know important but it MCB is nice that it can run both right so there's a whole bunch of different uh I guess there's a video as well so there this would recommend maybe watching that some of the Reddit comments overall I just think that it is uh interesting to think about why we need an MCP I am not like I he Frank might even

mention this i don't think that MCP is the perfect spec but at least it is under constant uh it's constantly being improved there's new drafts there's new changes and I do think that it is better than just trying to make APIs work so trying to retrofit that all right so that is it for our AI news for

the day is there any if there's anything that you think we missed anything else you'd like to discuss please drop it in the chat leave a comment uh again this is live so whether you're watching on X LinkedIn YouTube wherever uh you can communicate you can let us know and the next thing I did want to talk about so if you were paying attention to some of uh what MRA was putting out last

week we introduced this new idea of a what we call master 101 so our master course and I'll pull that up and we'll we'll talk through it a little bit and this was something that we were kind of really excited about because we I guess maybe I'll I'll start with the problem we were trying to solve so we always wanted to you know

master is if you're not familiar with master is a AI agent framework for TypeScript we it allows you to build agents and agentic workflows using Typescript right and you can build a build it locally deploy it anywhere and you can more easily it kind of gives you that framework around like how to think

about building agents giving agents tools um connecting workflows all all those things that you would need if you're kind of building out an AI application so it it's really like meant to be a you know a piece of a developer's toolkit if they're trying to add agentic features to an existing app or build a new app kind of around

building an agent but one of the things that we consistently ran into is we wanted to have some kind of course for people to get started because there are some new concepts and new things you know a lot of people have backgrounds similar to others on the Master team where we were from the web dev or the

app development world and you know we're not machine learning engineers we're not research you know data scientists we are you know we're builders we like to think of ourselves as builders we build websites we build web apps we build mobile apps like that's a lot of our background on the team and so often like you know we went through the process of having to learn a lot of these terms a lot of these new

things and MRO tries to make it simple by giving you like this set of primitives but it is uh still a lot of information to digest so we wanted to have this course that would allow you to uh more easily get access to kind of like getting this information really quickly but one of the challenges with courses especially when you're working on a framework like what we're

working on where things are changing pretty quick is that the videos if if it's a standard video course it gets outdated really fast um and so of course you can build a bunch of guides but the problem with like written guides is not everyone wants to sit and read through the guide and follow step by step some

people enjoy that but not everybody some people want to watch videos some people uh everyone learns in different ways and so we were thinking of what's what's a different way that we could build this course that would allow us to um really give people a lot of information relatively quickly but they could go at their own pace and then they could kind of get all the basics of MSRA pretty

pretty quickly so we we kind of came around this idea of we're calling it MRA 101 it's learn learning how to build agents with master and you know there is a video you know you can watch this video here welcome to the master course but uh it's not really a video course so you have an intro video there are three lessons there's more lessons to come but each lesson has

has kind of a it's an intro video as well and we can kind of go through the lessons so the first lesson talks about building your first agent talks about Maestra what you know has a whole bunch of different lessons here then we talk about tools in MCP you can see there's a bunch of lessons there and then we talk

about memory so a whole bunch of different lessons here and the idea is though rather than being a video course this is an MC this is a course as an MCP server and what this means is we can actually run this in cursor or in windsurf or in VS code and it will walk us through interactively the course so rather than me you know in this case teaching the course your

coding agent in your IDE or your editor is actually teaching you the course so this is pretty cool because it allows you to you're kind of in it's in context learning right you're actually the agent can write the code for you it can explain it it can tell you the code to write there's you kind of can

communicate with the agent like you're communicating with a teacher so rather than just watching a video and having to go through everyone going through the same learning path you're actually uh learning kind of on the spot as you're like actually writing the code so it validates along the way that you're

actually doing what uh what each lesson says so we do want to spend a little time we have you know maybe 15 minutes till our guests get guest gets here so we're going to spend a little time going through this and then we will uh kind of show how the course works and maybe if we have a little time we'll we'll talk through the code because the

code is open source it's available on uh on the master GitHub and you can our hope is that others will build more interactive learning experiences like this we do think this is kind of the future of education so curious what you all think um do have you tried this yet what do you think of it do you think it would be helpful in learning how to use any new technology not just MRA but you

know any new programming language any new tools if you could actually learn in editor rather than having to you know watch a course and then try to copy it or follow a bunch of written guides so let's go ahead and let's just take a look at the course in general and see how it works so I'm going to share my screen and we are going to just dive right in

all right so we're going to start with uh give me one second to just shuffle around some some windows here all right so we're going to start with just copying in and I'm going to use cursor but you could use windsurf you could use vs code i will say that you know one of the challenges of building a course in this way is that it

doesn't work the same in every uh editor right because each editor has a slightly different way that they've built their a their code agent and so I do think based on our testing it seems to work the best right now in cursor uh wind surf it works pretty well but it gives you a lot of extra information so like it shows the results of all the tools and it just it gets a little it's

a little messy the output's a little messy vs Code seems to be we haven't tested as extensively seems to still work but I would say based on my experience cursor seems to be the best and so let's go ahead and we're just going to copy this and I'm gonna go ahead and copy this code this is going to uh you know we need to go ahead and follow these

instructions but let's go ahead and just try it out so we're going to do mpm create master at latest and if you do try this there has been a few issues with just like the way that npm works and how the latest tag works so if you are having issues just try removing the at latest and that should also that should fix it if you are running into problems uh let's call

it master course june 9th okay so this is going to create us a new master project and the nice thing about this is it kind of just sets everything up for us so then we'll open it up in cursor and we should be able to kick off the actual course so it's going to take a second here but if you are tuning in in the chat let me know have you tried the master course yet what do you think of

this idea of uh you know a course as an MCP server so we're going to just follow through the defaults here and we will select OpenAI i will skip and just copy in an API key we'll add an example and then I am going to install cursor because that's what I'm using okay so going to go here and then I'm going

to open this up in cursor zoom in a little bit it says new MCP server detected so I'm going to make sure that this looks the same you can see here if I look at what it's asking me that looks right so that's what I want it to do and again I might have to remove this at latest we've been having a few little bugs with that but I'm just

going to save this it's enabled i am going to Copy in an environment file we're going to make sure going to delete this one here because I only need the one all right so let's go ahead and you can see we have agents tools workflows I can just now if we look at the documentation if we were to watch you know kind of watch the video what it'll tell you to do is basically just tell

say to start the master course so if I watch this video that's all I need to do is tell it to start the master course uh one of the things if you do run into problems check out the FAQs we're we're going to be adding more um I think MPX has a known caching issue so you might we might need to do that but let's test this

out so we're going to go into cursor we're going to say start MRA course and I think I might already have a course started create a script to start the course that's not based on the code you set up the master instance but there's no entry point or script to actually start or run the course okay well that isn't technically

correct i wonder if it's not view course status should be sending my uh should be giving me running these master tools and maybe it's failing because like I said there's that little bit of a caching issue all right so it's not quite working um what tools do you have access to so it feels like this isn't quite working for some reason

why i don't know let's go ahead and look here mcp tools all right so Mr says zero tools enabled why yep so we're going to try not doing this we'll just try this and let's see if this works i'm going turn these other ones off nine tools enabled i don't know why it's still showing red but at least now we have the tools enabled so step number one make sure the MCP server is actually

working now let's try view course master course status all right so here now you can see it's actually calling the tool so it says it's not been started great okay start master course okay so it's going to ask me for my email i just want to uh test this here so I'm just going to give it my email address with a tag so I know and it's going to run this tool so

I'm going to tell it to run it okay so it says you're now enrolled in the master course welcome to the first step tells me what an agent is tells me agents are AI systems that can do this the best agents use several important features like memory planning tool use and feedback loops all right so we can just

say when you're ready move on to the next step okay so I learned some things uh but let's go ahead and say view course status while we're here you're currently on the first lesson i don't think this is technically right as you can see we do call this course experimental for a reason because it is uh definitely very experimental that's not right

um because it should be giving me a unique URL for for my specific uh email so maybe I'll need to look and see oh there it is okay so you can see it probably did return the right Oh so it didn't call the it didn't call the tool here so you see I told it to view the course status it didn't actually call the tool because it knew where I was at the agent tried to be a

little too smart and when I did when it did actually call the tool it gave me this which you know I can see that right here so let's go and open this let's actually open in the wrong tab so if I go back now you can see I'm building a first agent here's where I'm at that is right you can see I'm on I can see the other steps you can see this is where I

am currently so let's go go to next step so this will move me on to the next step it says you've completed the first step what's MRA tell us a little bit about MRA what it has can build test and deploy AI agents and there's my core status should have just kept going next step and we can see that I should be able to refresh this page and I should you can see I'm moving through the

course right so that that is the one nice thing is and actually anyone could go to this URL i could share this URL with with anybody or open it up on any browser because there isn't there isn't really any sensitive information here right it's just this user's status of where they're at in the course but it is

nice that as I go through the course in my editor I'm able to see and keep track of my status so I can see where I'm at it's kind of provides you that uh ability to to go through it step by step okay so let's go ahead and say or let's see so verifying your master installation we should already have that we added our API key so let's go to

next step and yes uh thanks i think this is Tyler thanks Tyler we got it figured out but yes I had to definitely had to remove the latest from the args okay so I'll run the tool so now it's going to verify my project structure so you can see you know we do kind of step through pretty slowly but it is pretty thorough because

it's telling us a lot about Maestra if you were to read through this there's a source master directory that contains an index that's where I'm at agents directory and this one's a little bit different so because we we did actually recently change this so this will be updated but it talks about you know the agents folder is where agents are

defined tools and workflows this doesn't actually cover workflows yet that'll be the next lesson i do also like once it figured it out it does tell me the course status URL every time that's kind of nice next step eventually we'll get to writing some code the first lesson is a lot of just like information and confirming that everything's set up

right we need to run the master playground so okay I'll listen i'll go here mpm rundev and we will check that out and you can see that as you know as we are going through this It is basically kind of updating the status as we go so here we are you know the playground's running i could test this what's the weather in Sou Falls

which is where I'm currently at it calls the weather tool and it gives me the weather and if we now are go back here and look you can see that it talks about the playground i can tell it to go to the next step and it talks about understanding system prompts gives it tips on what makes a good system prompt and if we go back and and check you can see that we

are still making our way through the course so you know I did get you know it there are the first like six steps you have to go through before you really start writing code but it's a lot of good background information around um you know just getting the basics understanding of what master is and then you uh you're able to eventually get

into it so let's go ahead and say next step one and we're about didn't have quite as much time to dig into the code of how this works but the good thing is this is all open source it's all in the master repo so let's create our agent so it's going to try to create a financial agent so uh I'm going to ask it can you create the agent for

me see if it can do that so it tells me what I would need to do but I actually want it to do it for me so here we go financial agent has been created for you i could look at it and verify it here it is so you can see um you know it's way more interactive right i'm it's basically teaching me but it also can write the code for me it can tell me what it uh actually did and then throughout the whole process I just can

continually update and see that I'm progressing through the course and it if I finished this lesson I could you know go to this next lesson i could tell it to jump to the next lesson now I can kind of bounce back and forth but overall it just gives me a really good uh kind of handholding through all the

basic primitives of Mastra and I guess let let us know what you think of this so if you're in the chat if you're watching this live if you're just tuning in we have been spent a little time talking about what we think you know we could be wrong but we think this is the first course as an MCP server that we've ever seen so uh curious how how it's

been working for you all that have tried it we released it last week again just go to basture.ai and it'll walk you through all the instructions of how to get started and you can try it out learn Maestra also see if this uh if this is a learning style you like rather than just watching videos or following guides if you want in editor learning inside

cursor or windsurf or VS code all right and with that let's move on to kind of the next st next segment here of of our live stream so for this wanted to bring in a special guest so we are we're going to bring in Lauren here from Ampersand and Lauren welcome hey Shane good to be here yeah yeah it's nice to see you uh I'm excited because I know we

are going to be talking about how we can give agents access to even more tools which That's right which is great because anytime I mean that's one of the primary things that we get asked questions about is once people understand how agents work and how maybe MCP works but then they're also wondering okay well how where do I find

the right MCP servers what kind of how do I give my agent that I'm building access to more tools and more functionality and that's what we get to talk about today so I'm excited likewise so maybe it'd be good can you want to give a quick intro to uh yourself and maybe like what you do at Amperand yeah definitely hey everyone uh at live and

watching this later my name is Lauren i'm the co-founder and CTO of Amperand we are a company that helps agent builders AI builders SAS builders connect their apps with B2B integrations uh prior to this I was a tech lead on the Firebase team at Google so if you've used Firebase before you know that is a developer platform for building web and

mobile apps i was uh the lead of cloud functions which was our serverless offering our AWS Lambda competitor uh as well as uh I led the launch of a new product called Firebase extensions which as the name suggests it was all about connecting applications with uh integrations and other things to extend the capabilities uh so integrations uh

connections uh providing external context uh has been something I've been working on for a long time so excited to dive into that a little bit more today yeah absolutely and yeah I'm familiar with ampersand a little bit because I think that you know we in the past we had been evaluating a bunch of different tools for just connecting a bunch of

integrations and you've been doing it for quite a while now so you have you have a large amount of of integrations and tools available um I I guess yeah do we do we want to show talk about the amperand MCP server i know that that's something that you kind of recently been been launching or where would you like

to go with this yeah definitely i'm going to start screen sharing if that's okay yeah go ahead uh so yeah today is a a very exciting day for Amperan because we have a couple of launches uh so if you head on over to withampers.com go to our blog you can read all about it so we have three launches today one is our AIS SDK which

I'll demo in a moment our MCP server and our headless UI so AISDK and MCP server uh are both open sourced uh headless UI is al also open source all three things are open source and uh if you go to GitHub go to um github.com/amplabai you'll see our full repo here um as well as links to the docs where you can learn a little bit

more so uh at a high level what these tools uh help you with is get access to all of our integrations so we have about 150 connectors for connecting your AI agent to B2B applications um so if you're building an agent and especially in the go to market space where you need to connect with your customers CRM customer support tool uh what have you

instead of building those in-house uh you can use amperand now to get access to over 150 of those uh applications so we've actually built a demo app uh using the Ampersan AI SDK and uh Mastra to showcase a um a use case where if you're getting support tickets you want an AI agent to quickly summarize it prioritize it so your team

can deal with it uh we call this app Case Sherpa and I will quickly walk you through what the code looks like so if you are familiar with Mastra you're probably very familiar with how to set up an agent and Lauren can you give us a one click a zoom maybe maybe two clicks that way for those that are maybe looking at a mobile device or something they can see it that looks good okay awesome thanks Shane um so

you'll see here that we are creating an agent uh like you're used to except in this case the tool that we've defined uh is actually imported from the amperand AI SDK's MRA adapter so um here we have a tool called update record so this is one of several tools we offer and if you're curious what all of them are you can head on over to our documentation page and we have a list of tools here

for interacting with your uh integration uh as well as checking in on the status of those integrations and starting OOTH flows um so uh feel free to check those out afterwards so this agent um has one tool which is to update records inside um of the customer's uh CRM and for today's demo we're going to be using Salesforce so here is the prompt that we

give to our agent uh we tell it that it is an expert case triage agent who will analyze Salesforce cases to determine their severity and provide concise summaries we provide the relevant uh amperand environment variables we give it some instructions as to how best to review the case how to uh classify it

how to generate a summary and then some guidelines on uh how it it should communicate as well as how it should uh update our record using the update record tool provided by the amperand AI SDK so this is our agent pretty straightforward uh and in the same repo you'll also see uh a folder called uh AM

so this is a YAML file that has defined uh the integration so every Amperan integration is defined in a YAML syntax like this where you are telling us what uh tools you want to integrate with so in this case it's Salesforce and what you want to do so here we are um subscribing to cases that have uh been created and when there is a case we want

to give it the ID the subject and the description so our agent is actually going to be a web hook listener that will listen to uh the subscribe actions that are coming from amperand um if you've never worked with Salesforce before um you might not know this but Salesforce doesn't actually have a web hook API out of the out of the box so um

normally if you're building an integration yourself you can't actually get updates uh in real time you'll have to do um maybe a bulk batched um pulling um type of a read um for for a particular time interval um and we in uh uh at Amperan have actually come up with a way to get real-time subscriptions from Salesforce so um it's pretty cool

when you can get it by just writing a YAML file uh like this yeah I know we we've previously at Maestro you know for those of you that have followed the story we were building at one point before Ma trying to build a CRM and we spent a lot of time dealing with Salesforce and understand that there are

some complexities we'll say just you know little complexities with dealing with their data and their API yeah well okay so I'm preaching to the choir here you know all me I don't know about the rest of those watching live i guess let us know in the chat if you uh if you have uh played around with Salesforce API or and how what you think of it

uh so let's see it in action so if you haven't used Salesforce before it looks something like this um this is the uh cases view i get to see all of my open cases so we're going to go ahead and um add a new case for agent to process so um everything is in French on our website we don't know why but we've

noticed that our entire instance in the platform is displaying everything in French dashboard settings error messages we didn't make any language changes on our end this is affecting all users at our um fake company Flex Fleet uh including new signings and incognito blocking our team from using the um the product effectively so let us go ahead

and oh I forgot to include origin we got this through email so let's go ahead and create that case so what's going to happen now is this will um uh go through amperand we're going to catch this uh new case we will send it to our agent uh which for demo purposes we've deployed on railway and then this agent which is a master

agent that we've built will then use the amperan AISDK update record tool to now summarize our case so let's see uh our agent has now come back with a summary it's prioritized it as medium which I maybe maybe it's medium because the French population can still use it uh so it is now saying that the entire platform is displaying in French affecting all Flex Fleet users and

blocking effective product use despite no language changes so uh that was pretty cool we've uh instantly uh gotten the update from Salesforce and our agent was able to go and process this so now our human support team can go ahead and triage that awesome so essentially the way it works is you know through your web hook magic it triggers the web hook which calls the

agent and then the agent will then create the summary add a severity and then just go ahead and update this case automatically yep you got it awesome oh sorry sorry go ahead no no go ahead uh so that was our AI SDK and all of the tools that we've exposed are also exposed in our MCP server um so as a as

a master builder you basically have two ways of interacting with amperand the first is um what I just showed uh here in this uh in this code snippet right here you can directly import from our AI SDK the different tools we expose alternatively the other thing you could do is use our MCP server um and there's actually uh already documentation on the Mastra docs

under tools and MCP if you flip on over to the Amperand tab you'll see that um this is how you can add the Amperand MCP server can you give us a Zoom a Zoom on this one too yes awesome thank you yeah and so okay if I use the MCP server then how do I because I'm assuming there's a way then to configure what tools actually show up in that MCP server or does that get you know how do

I then if I only want it to have access to Salesforce for instance do I get all the tools by default is there a way to specify that how does that work yeah so currently you do get access to all the tools by default um but this is a great feature request that we will address in our very next update well I mean I I do think also though if you were

to you know call on this you were to call await tools you get the list and you could then selectively only pass in the ones you want to the agent so like you So I think there's there's there's ways around it there's easy workarounds but maybe that you know maybe there are uh other ways to make it easier but I

think that I I don't think we're necessarily uh there's pretty easy ways to control what tools your agents get but it it is pretty cool so if I So by doing this then I get access to all the tools that Amperand provides that's correct uh and then using the AI SDK you do actually get to pick which tools you want so you'd only import the tools here that you want your agent to get access

to um and uh you would put them in this block right here awesome so you have a little bit of uh flexibility when you're building out your agent if you want to give your agent additional capabilities connect to different services you have a few different ways to do it depending on what you know how you're architecting your app or how you want it to actually work yeah definitely

very cool i for anyone that's in the chat watching this uh let let us know what you think let us know what kind of uh tools you want your agents to have i'm sure Amperand already has them you can just uh wire it up and be off and running yeah definitely so um for those that are watching uh there's a few few things um where we you can get involved and we would love to hear from you the first is

if you go to our GitHub you'll see uh a number of um open- source repos uh the the two that are the most exciting um is one our connectors repo so we've open sourced all of our connectors um and you can browse this uh here for how we're calling the various APIs and uh there's some more information here you're also

welcome to contribute a new connector if there's uh one that you really want for your agent uh but we don't have it yet you can uh make a pull request you can file issues here if there's any um uh that you want to request for and the other uh repo that I will uh call your attention to is our AI repo so this is one that we um uh that we uh just uh

launched today so this has the source code for both our MCP server uh as well as our AISDK and also welcome uh to file issues make contributions we love to hear from you very cool yes so and and for those of you that know you know anytime we show these open source projects we always say go and give them a star you know go get on go on GitHub give them a star while you're there give Master a star if you

haven't already you know we we do don't mind having a few extra stars getting thrown around on GitHub yes awesome it's very cool that all this is open source too i think that's uh it's always nice to see and I think you know having the connectors open source makes a lot of sense because a lot of times there are here some there obviously you have all

the popular ones right like any of them that are common you already have and you've you know are very well supported but there are probably some that are slightly less common but still maybe very important to us you know some groups of people and the ability for you to either request them or you know maybe even try it you take a crack at starting

to implement them is is a great thing so awesome is there So Lauren is there anything else that we wanted to cover today uh the the the very last thing that I I'll talk about is uh a little unrelated to building an agent but I think still pretty helpful to most agent builders um and that is the headless UI library that we launched today so um up

until this point we uh have provided a UI library that will um that has pre-built components that you can directly embed into your application and this will um help your user to both connect their SAS application as well as configure the mapping and optional field selection um so this was really convenient and easy for for builders that want to get up and running but we got a lot of feedback that um sometimes

people want to customize the UI uh and while we do offer some uh basic styling on our pre-built components like customizing the colors and the fonts um you couldn't for example switch the order of the pages or or completely redo how the layout looks so what we also shipped today is the headless UI library so uh this is uh a part of our uh

amperand react mpm package uh and it gives you all of the hooks and methods that you need to build your own UI so uh this uh helps you with the state management so um a few things that it helps with is uh managing all of the connections so knowing if a particular user has already connected to their SAS application um or if you need to guide them through

a new connection flow uh you can also manage installations so installation is our terminology for a connection that has configuration also uh so so this is the unit uh of of a um integration that a customer is dealing with is called an installation so uh you can manage that uh you can even get um some schema

introspection so um you might know that a lot of CRM and other tools have the ability to allow people to create custom objects and custom fields and those um you don't know about until the customer has connected their tool so using our headless UI you can also actually do a real-time scheme introspection into all of the fields that exist in that customer's object um for example what

are all of the fields on their CRM account object and then you can build UI uh and logic around it we also help you do things like managing the uh reading um configuration so this is uh any field mappings that your customer wants to set up where they may have mapped a custom field to uh a particular concept and

then once they uh set the field mapping we then in um in our reading and writing we'll normalize the data and send it to your application in a mapped way so your application gets nice clean data to deal with and then we're dealing with all of the uh complexity all of the mapping all of the keeping track of all of the

different schemas underneath the hood so that was already available in our pre-built uh UI component now you can do that yourself uh in our headless UI also and then uh writing back we also um have a lot of powerful features here so when you're writing back to a particular particular uh SAS application um like we

did in in the case Sherper demo we wrote back uh an update to the Salesforce case you can actually set default values so in our app we could have had the severity of the case set to let's say medium by default so if we don't specify severity when we're writing back it is just set to medium by default

you can set integer values boolean values um you can also configure what happens if your customer already has a particular field set so let's say that um a human actually went in uh that case uh before our AI agent was able to summarize it maybe the human was lightning fast caught that uh case right away already categorized it and already

summarized it um we could then set this write on update setting to never it means if something's already in that field we're not going to overwrite it um so that's another way you can configure the uh the the write back and you can put it all uh together we um also open sourced a demo app that showcases how you can build your own UI

so that's uh again available on GitHub demo- headless um also open- source MIT license welcome contributions and issues and um yeah and that's what I wanted to show today so three exciting launches headless AISDK MCP we love your GitHub stars yes yes so let's send some GitHub stars their way uh yeah that's really

cool and and even with the headless UI library I know that you know if you're building an application where you need your users to authenticate to a service and then you want to have that connection and then you want to use that whether you're building an AI agent or you just need that data from other some other purpose i can uh I can see where

you know ju using the pre-built gets you farther faster but ultimately you might want the flexibility to being able to say like okay I want to customize this to match my design to match my you know my style right of whatever the app is you're building so that's pretty cool thank Thank you yeah we we definitely

found that in general where customers when they're first starting nail to build an integration um they want to do something quickly get to market um and then as they're in market as they get bigger customers especially mid-market and enterprise customers is when they find they'll have to think through um uh

how to how to make things more uh customized how to own the user experience more how to take advantage of some of some of our advanced features like field mapping uh all of the advanced uh right options that I talked about so it's definitely uh a journey where uh it starts out simple but integrations can get complicated pretty

quickly and yes as I'm sure you know building out the last CRM product yeah integrations are you know they get messy yes definitely uh and our mission at Ampersan is to really help developers along that whole journey to get them started and and get their product to market really quickly but also be with them as they are scaling as they are going up market as they are uh really

expanding out their integration depth yeah and and we have a lot of people watching this that are building AI tools AI agents and I think anything that we can do to help you know help those people get connect to different services more quickly yeah it's allow them to get farther a lot faster so this is you know

tools like this are really helpful yes definitely um yeah thanks so much for having me this has been really fun and uh yeah uh please uh check us out and we have a Discord channel also sort of Discord server yeah what's the best is that the best way to keep up with what's going on with with Amperand yes definitely so um so if you go to uh with

ampersand.com under this developer tab there's a link to join our Discord community awesome so yeah please uh if you're interested in giving your agents superpowers with a whole bunch of different tools at once uh go to the Discord keep up with what's what they're all shipping at Ampersand so thanks Lauren it was really great having you on talking about all the launches a lot of

cool stuff coming out and yeah if you are are building a MRA agent it's already in the docs just go to the MCP overview page in the docs you can see that you can connect to Ampersand's uh MCP client very easily and get access to a whole bunch of tools as Lauren also showed there's other ways if you want to more directly control the tools that

you're sending your agents you can use their AI SDK and just pass in the tools that way thanks so much for having me today Shane yeah thanks Lauren we'll talk to you soon yeah talk to you soon bye see you all right everybody uh that was a good solid AI agents hour we talked some AI news we talked about the Maestro 101 course which I think is the world's

first course as an MCP server um so check that go back and watch the recording on YouTube if you want to see that otherwise just go to master.ai/course and try it out for yourself and we talked with Lauren from Ampersand around uh how they're making it much easier for you to pass a whole bunch of different tools into your agents and kind of giving your agents

access to a whole bunch of different services uh really easily so thank you all for tuning in and we will see you next time on AI Agents Hour goodbye