AI Agents, MCP & Blockchain: The New Interaction Paradigm - Damjan Dimitrov | Apillon

Hello everyone, my name is Damian. I'm here to talk to you today about AI agents, the MCP protocol, how they converge together with the blockchain technology to form a new interaction paradigm. So uh as a bit of forward appilan is a web tree development platform where we build different kinds of tools and services for to make it to make developers lives easier where they want to build deps or some other solutions and uh facilitated their whole work and one of our recent initiatives was our own MCP server which we're going to dive more into in a bit. So first of all uh let's define what are AI agents. I'm sure all of you here have already interacted with some kind of LLM before.

Whether that's chat GPT or CL bionthropic or Google Gemini and so on. We all know how uh increasingly better they become at generating responses or even images now giving us some some text uh summarizing data, processing data and so on. But what happens when we needed need them to perform some actions for us? Like what happens if I come back home after this conference and I say hey can chad can you download all the images from mil is billgrade for me well of course they don't have the capability to do that but as soon as they have some uh capability to perform some actions for us even in an automated way that's the moment they become actually an AI agent and the word agent also be uh comes from the nature that they can act autonomously so they have all the context injected to them they are aware of the environment they're and they have the context and they can operate autonomously uh as we give them as as we program them to. So they're meant to be nondeterministic meaning that every single time they're requested to do something they're going to find the most optimal path to perform that action for you given that context.

Uh with that said, they also uh have are able to learn and adapt from the context because they have uh knowledge of the previous actions they performed and the previous outputs they've given you. So they're able to continuously uh refine the outputs they give you. And finally, they're meant to operate across platforms uh across uh any different kinds of environment. And in our case, that's also works for any different kind of blockchain that you connect them to. So um the way they work is here's how I like to think of it.

If an LLM is the brain uh so it has all the knowledge it has it the context it's uh it's aware of its environment where it's operating in and it has all the context. Then the moment you give it some capability to execute some actions for you that's when you give it hands. So you connect the brain to the hands and now it's able to do some actions for you. We're going to dive uh deeper into how we are able to uh give some capabilities to an AI agent. But you can remember that their core loop is just to observe and then plan based on the context it has then invoke the tools it has access to and finally they give you the results and refine it over time.

Um let's see why blockchain the as a technology and even DeFi is an ideal spot to connect with AI agents. So let's uh cover some of the most uh basic properties of of the blockchains. They have they operate within a permissionless environment. So nobody needs any permission in order to just spin up a wallet, generate a wallet, deposit funds to it and perform transactions on chain. Uh then it's fully transparent meaning that any action and transaction that happens on chain it's uh there forever.

It's completely uh irrevocable and anybody can verify that it has happened at a given point in time. Uh it's meant to be fully programmable especially with smart contracts like on Ethereum where we can define some complex logic in in order to be executed automatically or on demand. And finally it's a completely trustless environment and gives you the security based on all the nodes that are securing the chain around the whole globe. So the point here is that all of these properties that a blockchain has inherently are the direct benefits to AI agents to operate on top of them because an AI agent if they need a wallet and they want to transact on any chain they can just instantly generate a wallet and once you deposit some funds into it, it can immediately interact with uh any blockchain that it's operating on. All the actions it it performs are completely transparent and you don't need to trust that this has something happened or something was executed but instead you can verify on chain.

And just like the smart contracts are meant to be completely programmable and they they even operate good with smart contracts because they already have all the pre-built data that a smart contract can uh have in the form of input and output. And of course uh we don't need to rely on trusting the underlying infrastructure because it's all running onchain and it's all secured by all the nodes. So whatever the agent is interacting with on chain is going to be completely trustless and secure. Let's explore some cryptonative use cases just to give you an example here. Um let's say you have some autonomous wallets.

So we all know that when you're exploring a new environment whether that's a new chain or something similar then you there's some um uh barrier of entry when you need to adapt to a certain wallet you need to research what are the best ones or what you want to adapt to at the UI and so on. Well if you have an AI agent which can connect to an external wallet then it can be able to perform these actions for you. It's able to for example do swaps for you transfers for you just by using natural language. you can explain to it which action you want to do instead of having to go around and adapt to the UI of a wallet. That's a pretty basic example.

Then you have smart contract co-pilots. So over time as uh coding co-pilots become increasingly better, they may help you identify the things in smart contracts that uh you you've probably missed or can help you debug or optimize your code. We know that in most cases where there's an exploit on a smart contract, the reason is that you've missed some kind of validation or authorization in the smart contract and AI can help you identify which are the points of validation that you've missed in the smart contract. Then we have uh DeFi assistance. So we all know there's some kind of research step involved when you want to interact with a new D5 protocol.

you want to learn which are the most reputable protocols, the most audited ones, the ones with the highest TVL and so on. So instead of you going and doing this research, you can just have an AI agent which is um which has the capability to do all the research for you on any different chain that you give it the ability to. So you can tell it fun me all the protocols which are fully audited and have at least this much DVL. Or even as a further step, you can tell it here's my here's my USDC. I want to lend it to earn some yield.

Find the best protocol and deposited there uh just so I can earn yield passively. Um then we have the delegate agents. So similarly as before, instead of you if you're involved with governance, instead of you having to do all the read all go through all the proposals and read them and analyze them, you can have an AI agent which gives you the whole summary of prop of recent proposals. It can find you the most reputable people to delegate your votes to or even it can vote on your behalf. If you want to if you want to completely have it as an automated process once you give it some criteria about proposal that you want to vote on then you have onchain researchers which is a combination of uh similar related to correlated to all the previous examples.

If you want to analyze some onchain trends or onchain data you can have an AI agent which can automate this whole process for you. And finally, probably one of the most uh uh frequent use cases are social and trading bots just because they're kind of a lowhanging fruit. There's many launchpads and frameworks which already have this built in out of the box with capability for NAI agent to post let's say post on X on your behalf once you give it some persona or it's able to analyze um on train macro trends and and uh perform trades on your behalf for example. So we've explored we we've explored some uh use cases which can help anybody in general which is in crypto and wants to interact with AI agents. But what about some what about the how it can help developers and and facilitate the developers work.

So for those of you who are uh in coding here and uh have used AI for as an assistant for coding before, you know how uh helpful it can be in speeding up your work and easily becoming to a prototype version of your project. And increasingly as we reach a stage in crypto where we don't treat projects as just projects anymore, but perhaps also startups, then we know how crucial it is for them to reach the validation phase and get feedback as fast as possible. And with AI agents, this can uh be brought to life much much faster. The time you you reach to actually get validation and reach the MVP phase of your product. They can help you debug your codes, optimize your code, um reduce bugs and improve security.

And of course, they can help with automation. So any kind of work you have which is let's say too manual or too repetitive then you can automate this using the AI agents too. Let's analyze now. Let's look at look look over the MCP protocol. So by definition it's a protocol which connects uh an a client or an LLM with some external data source and tool.

So if I have some uh LLM like for example chat GPT or cloud and if I have some external resource then through the MCP protocol I can easily just plug in this data and it immediately has access either to some external data and resource or gives it some capabilities to perform some actions. So if you go back to the previous example I gave, I come back home, I I tell Chad GPT to download the photos from ETH Belgrade and of course it's not going to be able to do that. But let's say this is on some storage service which has an MCP server uh already created and then I can very easily just by giving it the URL I can plug in this MCP server to my AI agent and now it instantly has access to all the photos and I can just tell it what to do with them. And even further, if I want to uh include some uh image filtering or image processing here, let's say I want to I wanted to find all the photos between 10:00 a.m.

and 10:30 a.m. just because I want to see the photos of my talk right now, of course. And um I tell it to do that and then I can find some another MCP server which allows an AI agent to have the capability of image filtering and processing. And then it can combine these tools together in order to bring me uh to the action I wanted to execute finally.

So why does it matter? U for those of you who are developers here, you know how the rest API schema actually facilitated developers work. You instantly see all the endpoints uh you have on your on your disposal. You see which data you're able to retrieve, which data you're able to post and modify on the server, and you see which are the inputs and outputs for each action. Well, this also works for AI agents.

You but the thing is that every time you want to integrate some external API, you need to give it the documentation. You need to explain each endpoint separately. You need to give it more examples of some inputs and outputs. And this is not completely agent friendly. It's not really AI native.

While on the other hand, MCP server is built to be AI native. all the all the actions and tools inputs and outputs and results are explained in completely natural language and it's meant to be in a way that's completely uh agent friendly. So this also facilitates developers work because if I have some external data source I want to connect to my AI agent. I don't need to do all the manual work myself. I don't need to code it and implement the tools but I can just find an MCP server which does this for me and I can connect it.

And it's a protocol because it's also completely standardized. So this can work uh um regardless of which AI agent it is or which environment or blockchain it's running on. Um I'm just going to quickly tell you about our own MCP server which we built recently. So as I mentioned before we have different kinds of uh uh tools and services. So for example the centralized storage hosting smart contract deployment etc.

basically tools which allow developers to more easily build DAPs and it's meant to be completely chain agnostic across uh any major substrate and EVM chain and it's completely plug-andplay which means you can just connect the the server URL to your agent and it immediately gains the capabilities to run on top of these services and now I have just a quick recording I want to show you so it's demo time just you can get a visual idea of how this demo might work so if you could please cue the video. Um, I have the cloud agent here, uh, cloud bionthropic and I ask it to I it's already the MCP server is already connected to it here. And now I ask it to, uh, list all my storage buckets. I think the playback is a bit slower here. Um, okay.

It's a bit slower. Uh, but doesn't matter. I'll just go through it quickly. So I ask it to list all my buckets. So a bucket is like a a logical container uh which we use to store files on the decentralized storage.

We use we have our own IPFS gateway. So I ask it to list all my buckets and it tells me that I don't have any buckets in my appilon account and also ask me a follow-up question if I would like to uh create a new storage bucket here. And then um once I uh I think I skipped forward now a bit. Um okay well doesn't matter. I'll just summarize what happened here.

So I then ask it to create a new bucket just by using natural language and then I ask it at the top you can see I ask it to store all the files inside the files folder inside that bucket. and it recognized which actions it has access to. And what it did is didn't immediately just upload the files, but instead it went through to list all the directories, list the allowed directories it's able to operate on top of. And then once I s it recognized, it found actually the folder it's supposed to upload the files from. It went then it goes ahead and uploads all the files.

I think it froze now, but it's fine. I I guess you can uh see the general idea here that the point is that you can talk to an LLM just by using natural language. You can explain to it what you want it to do and by having the MCP server plugged in, it can immediately uh recognize all the capabilities it has, all the actions it has access to and it will always find the most optimal logical route it needs to take to execute that action. So it's not like oneoff request response it were with with an API but it can recognize which actions it needs to take exactly and which is the most optimal path. And furthermore uh if it runs into some problem or an error if it were an API you would have to debug the code yourself and analyze what's what what went wrong and then fix it.

But on the other hand with this it can analyze if it did some mistake on its own and then uh basically find a better path to do or ask you for a for a follow-up question and explain what went wrong in natural language. So we can go back. Yeah. Uh as a key takeaway here uh I would just like to say that we learned that AI agents unlock autonomous crypto experiences meaning you can facilitate and automate your work much easily much more easily. Uh the blockchain as a as a technology and as an infrastructure layer ensures you uh the AI agents run uh securely.

All the actions they run are completely secure and transparent and they run within a permissionless environment. The MCP protocol ensures uh safe integration between external tools and data sources with your AI agent. And of course, you can already start building this and experimenting uh today. Um that's that's it for the presentation. Uh if you have any questions, feel free.

Or if you want to see a more deeper deep dive into this, I can I can show you a demo be around here today and tomorrow. So you can scan also this QR code to learn more about Appilon. And that's all I had for you today. Thank you. Okay.

Um, thank you for the talk. Let's see if there are any questions now. Okay, I see a hand there. Um, thank you for your presentation. I have two question.

Uh first of all uh is it possible through your MCP server uh to do transactions and if so how do you handle private key sharing? Yes, it's it's uh possible to do uh transactions and the way it works that is uh you can run the the MCP server locally and the thing that happens is that you give it uh you have a list of environment variables that it has access to and this can run either completely locally in your machine so you never post the environment variables anywhere to the cloud or or you can uh for example run it inside a trustless execution environment. It's also possible to deploy the MCP server between a within a TE and then it runs in in a completely trustless environment. But also on Appilon the way it works that is that we abstract away the the whole blockchain layer and instead when you do transactions you do it with your own account and pay and pay just with credits instead of crypto. So the the point here is that we want to completely abstract away this whole layer and even allow web two developers who want to build on top of blockchain to come without having to spin up their own wallets.

So that's that's the way it works usually. Okay. Are there any other questions? There's a hand right there. Uh so imagine every user in DeFi has a DeFi AI agent to do like yield farming or investing.

You think that could lead to strange market effects?

Um perhaps perhaps uh it it depends. I think it depends if you because an AI agent can be some external tool that you use and you just make it uh make your own instance of it or it can be uh something that you completely program on your own. So I think it depends on how the AI agents will be tweaked in the future. For example, I can make an AI agent which does trades every second while anyone anyone else will uh make an AI agent which can perform transactions every day. So I think it's um but but even we've seen this even in let's say traditional infrastructure now most of the trades being executed are bots and this is noticeable but it doesn't have such an effect and maybe to to a certain extent it even brings a benefit because it increases always liquidity like with automated market maker for example.

So uh can't can't know for sure. depends on the way it's built to behave, but we it might be a situation where we notice some differences or it might be an obvious behavior that it's actually an AI agent instead of a human.

Thank you. Okay, do we have more questions? Okay, if that's it, let's please give Damian a big applause. Thank you.