Introduction: From MCPs to Agents

I want to kind of guys walk you through a little bit of a different use case so a lot of people here who here has used MCPs okay cool what do you guys what do you for those who are actually using MCPs what kind of how do you guys use

the MCPs anyone use one or just heard about it okay you're using it okay okay that's an interesting I'd ask you about that later so here who here has heard of agents yeah a couple anyone actually build them or anything like that or more

just that looks cool okay we're gonna get in the back I was okay cool so I'm

sure most people if I mention API they will probably know it much more what we

Shout‑out and Inspiration

wanted to do is actually what's I want to give a shout out to gentleman who actually organized this it's a gentleman named Rom old here later if you guys want I'll post it probably post event he has an MCP server he's actually probably

really cool for it because if anyone anyone here heard of n8n yeah people use it so what he's done is he's actually gone in his own time I thought it like I saw this guy and I saw this thing on YouTube and this gentleman had created

An MCP that Builds Agents

like a full MCP that can create agents for you so basically he took the full documentation set up all the documentation he took all the things that are involved with different type of what that build might look like and he

actually in his own time created this thing put on YouTube and it really took off and this is a little bit of a shout out to him because I think it's it's a

really useful tool so later if you guys want to ask me about it I'll show you guys what it is he's got his own website and stuff but uh i added my own spin on it so one

No-Code and Low-Code Automation Landscape

of the things that was good about it is a lot of people when they have has anyone here used any platforms besides n8n and local platforms to just drag and drop things okay lang chain is that's that's cling on to most people but I like any Dan make how about anyone here

Sapir yeah so what what is the biggest thing so I saw all of you probably started somewhere and we're like this is really hard right so what I'm going to show you is the way because everybody's probably seen the chat GPT drop we're

building agents I don't know I've tried to find in Europe apparently it's not not available but i've seen enough videos if anyone's used it please approach me after i'd love to learn how you got it set up but it's the future of where things are going they mentioned

Our Spin: MCP as a Tool Layer

some mcp servers some things that are going on this is our own spin because to give a little bit of background we've created this a little bit different because we've taken the server

which if you think about an api it's actually a tool so if you have an agent and if you used agents they're actually just a tool so every time you're setting up a tool you have to have an api

connection so there's a bunch of things you have to set up it's a very long process kind of painful you don't know if the tools are working as well as they could so what basically anthropic did

What Is the Model Context Protocol (MCP)?

last march and it's really taken off is created what they call model context protocol which is basically a toolbox so now ai and agents are able to actually figure out what system that you're

Demo Setup and Environment

using so that's what i'm going to show you guys today uh let me start off so i'm using an application just for for what you're looking at here this user interface uh thanks to my co -founder sergey here he'll speak later uh he helped create a user interface most people set up in cloud we

set it up differently because we set up in dust for a couple reasons so one of the reasons that i that i prefer this setup is because the way that we have it set up if you actually go into the

Agent Anatomy: Memory, Tools, Instructions

builder here the platform itself so agents are made up of three things memory tools and instructions or the brain or the llm right so if you look here the knowledge that it's looking for

Curated Knowledge and Troubleshooting

it's got information related that I added to it, that is the best type of NADN designs, the type of curated workflows or examples, readme files so it can actually go through those and try to find what would work best for that, as

well as troubleshooting and so it can actually fix its own and debug its own problems.

Instructions and Tooling via MCP

Of course if you look here there's also instructions, workflow engineer, there's basic instructions to look for, but as you notice here the

it's actually looking for mcps and of course later we use cloud for this but the main takeaway so if you look at the tools that we have set up here i made it have available this tool here which

Live Build: Audience Use‑Case

i will turn on just to show you how it fully works so anybody have an idea of an automation that they want to see built live. Anybody have any ideas?

Anything that really annoys you at work that you wish you don't have to do?

Sure. Okay. So...

Anything with content repurposing is usually a really good thing.

Oh, he talked to Sergey after. He's going to be brilliant.

Check out Gamma or really talk to Sergey, though, really talk to Sergey.

Look at my presentation in MyStand from July. Yep, yep.

Defining the Personal Assistant Agent

it so you want to have something that's essentially like a working brain of anything that you've done so why don't we do this so you want to create an agent that is essentially a remembers all your interactions yeah so like a personal assistant type agent okay so let me show you how this is done I don't

guarantee the results forever this is a starting point I always say to anybody who whatever they're building this is something that you have to look at and take into consideration that it's literally just a starting point but if you can iterate of course anything if you've got a prototype fast you can move

quicker so let me just put that in do you have any other any other requests or anything that you want to remember specifically anyone anyone summarize transcripts or something like that?

Desired Outcomes and Channels

So give me insights from the day and then let's say, give me some questions I should think about.

Let's just make it simple, right? So, okay.

Transcripts.

Yeah, so take my ideas and post them to my Telegram channel and all other places. All right, take my idea.

All right, let's try that. I don't promise amazing results, but we'll see how it goes.

Under the Hood: How MCP Orchestrates Tools

so once i click here it's going to give it some thought normally what you'll see here is when you're working in 8n i'll just open up so you can see how it actually works right so the model context protocol will actually be working here and as you're looking through it

what it's going to be able to do is it's a layer on top there's an llm layer on top of the server and it's looking at different things that you can try to use so it's got data

query you can see here the agent when I ask it to do something simple like this it's got all the instructions says chickens memory what it remembers when it's actually looking through here this when I was showing you earlier with the

Context Engineering: Tuning With Examples

setup what this is is basically the memory so I picked selective files do So what I've done is I've actually regulated,

Do you know what context engineering is?

think of it like tuning a guitar. I've tuned it to specific notes by providing these specific examples or things that it should look out for.

So good builds in N8n or any of these files that would be built and I've actually built it right into process. So now it's going through.

Error Handling and Best Practices

There's error handling built in.

So anything that would be things that you would be, let's think of it like a expert that has ability to build automations but has instructions what are best practices right rules instructions and examples

actually it's so part of it they'll be files that are workflows part of it will be readme files what those workflows do to add context and part of it will be best practices that you should and should not do when you're building automations right

Build Process: Iteration and Resilience

how much time it took to you to create this this particular layer because the mcp took longer what i'm adding is a layer on top of the mcp which essentially but it took me about three hours most of the work was one thinking about what sort of information

do I want to have for the agent well that's great on the exit see what happened here let me just say try again here's the law of demos right but

basically what I've done is what I was looking through all those files I added examples I added specific things what are good things to do so the workflow that I put into the prompt was you can check online or first you would check

here is what you have in the documentation the second step would be if you can't find an answer in the documentation look online and then the third if it's not something that makes sense according to these rules research

search it again right yeah yeah yeah I'm trying to walk through the example but it might have timed out but if it doesn't work what I can show you is kind of all the servers down okay not let explain yeah if you could that would be

great so just give me one sec because I think that's actually a big part of it can't really do a demo that's why it keeps reading as you can see it keeps keeps reading but basically look at this for example here if you go through the

Workflow Patterns: Loops, Batching, Throttling

efficient loops that it's looking through right so the way you're asking earlier the type of files that I'm looking at right there might be batching splits and loops bulk endpoints these type of things so kind of stuff you'd probably throw to a agent that you're working with that would be able to try

through and do that right so it might give instructions like when you get a lot coming at you files in an automation slow down take a break so you don't process everything at once so you don't time out the API for example right but

let's see let's try this it's gonna keep struggling struggling here let's see what it does keep doing so we did give it basic general instructions but

Planned Outputs and Routing

But here you can see what's cool about it as well is when it's done the output, it will actually give the instructions what it wants to do.

The input, so it comes from a webhook, so something's coming in. Process the content using GPT -4. Output some sort of code, so it could be JSON as well.

Context rooted to be in a certain way, so a social summary versus post summary. So kind of like what you guys were asking, you know, maybe looking at something more.

that looking at potentially some error handling that type of stuff and then generally because the MCP if it's not gonna work I'm not gonna take up much more time you're good to go all right so now that's working build it all right so

I'm just gonna try this and this will show you the yep sorry go ahead so any other question did I lose anybody yet no okay good so now that they see

Security Considerations and Reliability

that they've got the mcp and this is the thing about mcps i don't know has anybody here read the actual documentation so it's going to ask but has anybody here in the meantime heard anything about you know security issues with mcps has anybody read that that's not fair you don't

up and seen some of the stuff data leakage these type of things right so here it's asking me to use say that no has anybody heard about google summaries showing up people's private conversations from their chat gpt no yeah the shared ones right it could be shared ones not necessarily shared because chat gpt has all sorts of interesting approaches to data but but what you're seeing

here actually now that it's working thank god thanks for great for fixing it it's actually

Result: The Agent Builds the Workflow

going through and we came up with a personal assistant it's actually going through and it's building it so it normally a guy if anyone works with automations it's going through the steps it's actually building a workflow that

automates personal reflections plus content material potentially it's gone through it's gone all the files that it's looked through as you can see here

what's cool i don't know if you guys anyone have a preference for a certain uh model that they like to use for when they do stuff like this claude everybody loves claude i went back to chat gpt recently and it's not that bad it's actually pretty good yeah it's actually pretty good

Why Dust: Generated Artifacts and Usability

so check this out normally um so it it's even looking here like if you look at here when you you do the personal assistance it's going to show you it's been created and

why we use dust and why I liked it personally so it tells you what the name of it is and then it creates a full file how to use it so it gives you more

technical sort of things that will be created but the main goal here let me

Overview: From Broad Prompts to Connected APIs

just refresh this let's go to overview so we've given it very broad instructions most of the time you're going to be much more specific you'll get a tech spec most of the time but as you can see the content comes in it gets

processed it actually gives you details so you can hook it up whatever it will be using the API so all this stuff normally this stuff would take at least two days to build yourself let's say potentially it could be this is taking

15 minutes we know what I'm talking about whatever but it's basically created a workflow that you just need to add credentials to and

Takeaways

Wherever you want to plug this in you can change it or change it out, but essentially your first version of an automation Because of the MCP the agent actually creates itself by using tools to create what you're looking for So that's kind of it

Hope it wasn't too technical my first technical sort of demonstration, but um

Usage Guidance and Verification

hope you guys like it does anyone have any questions or anything about it or seems pretty straightforward well most of the time obviously I will always say it's like vibe coding you want to make sure to look at what the workflow is what I'm giving here is post to social media for

example right there's a router here there might be assumptions but I always recommend to look at the output what's going through what the workflow actually is and verifying yourself but if it works technically let's say you agree with it you could just set this up plug it in put in the credentials

Credentials and Running via MCP

so api keys whatever it might be and you're good to go you can actually ask an mcp server to run so if you already have credentials set up you're good to go like i'm very lazy person i do that my life's been a lot simpler since so yeah but yeah i mean i'd be happy to show you what

Effort Saved: First Iteration in Minutes

server it is later and stuff too anybody else have any so the general idea is that this book can create this flow without us interacting with actual yeah forget sitting there and figuring out what every node in n8n means because most people let's be honest i mean that's you you You know, you're doing automations, it's a long time to learn it, right?

You can get clawed through an MCP server now to give you a first iteration. And if anyone's built any automations, this basically has removed like two days of your work in a week.

Q&A: Tuning, Layers, and Practical Limits

Actually, you said that your tool was tuned by you, your upload device. The tools that I built? I'm using somebody else's MCP server that they created so I made my own layer

on top of it so I added specific context right so I made it my own personal business context so if somebody has a tool that they have and it's built and whatever it is that type of information is just a layer on top of a whole other

so it's not a wrapper because this is different because this is actually able able to give instructions to something to what you might have seen like you know chat GPT agent it's not web browsing but if you have small tasks

How the Context Was Chosen

the the UI think of it that way right oh good question the documents that I

picked the documents that I picked was based on experience on output and a few iterations based on the original MCP so the original MCP is very general it gave

me like you know the the LinkedIn workflows where you see like a hundred thousand whatever it is things that would break right so I decided that I think the biggest problem was you know for example instructions like look don't use more than nine nodes or nine workflows for the data pass through

certain things like that really stabilize what kind of output it would give so it was like first set it up see how it works run it through this user user interface see what comes back with come back with what went wrong find the documentation the

Continuous Improvement of the Agent

two hours that I said to set up this particular agent that you see here in dust and give it the proper documentation when I just continue add if I find something new something whenever I output something I always get it to give me that's why I give it the summary summaries here it gives me

everything I saw so it would look like for example error handling here right kind of gives a background of best practices but this is something that i will add and i may add different layers or update this so each time i run a new iteration of this agent to build this i'll have uh hopefully

On Credentials and Data Hygiene

better output and another question because here it's connection with the social media already you don't want to give AI any information yet this is where people go

so you think would be a good idea to put in a dust logs and the passwords or to do it by yourself

broke from credit cards stuff like that so don't but if you've got low -level use cases I'd be happy to talk to you about it more but there could be just by basic

Closing Thoughts

first saying authorization type levels right so you guys think good button