Apify, MCP & the Agent Economy: Jan Curn | The AI Rabbit Hole 2025 | DVC

Hello everybody. Let me see if this works. All right, cool. Okay, so the broader topic is agent monetization, but I like changed the subtopic of the presentation a little bit. and today I'm going to talk about how to enable AI agents to discover and pay for arbitrary services. I'll get to the motivation during the presentation but before I start I wanted to ask just out of curiosity how many people here know Appify? Well, it might be like 10%. Pretty good. All right. So a short introduction right started 10 years ago actually in YCO major fellowship and we started as a web scrapping company like a fullstack web scrapping platform where you can run scrapers and over time we actually launched a marketplace of these tools or scrapers actually we call them actors and now over the years we grew to 5,000 actors for all kinds of use cases right so for example we have a lot of scrapers still like for example if you need data from social media.

We have scrapers for all the major social media sites. We have scrapers for e-commerce sites, search engines, maps, all that. And we have also like a general purpose scrapers for example if you need data for training your AI models, right? So you just scroll website like Google does, right? And converted data to markdown to feed it into your like AI chatbots and rack pipelines and so on. And over time actually we started to like bring even like other actors to the store u including AI agents as well now and data processing tools and UIs and so on. So the ecosystem actually grew and some of these actors are built by us as a company some are built by by our community and just last month we paid about $250,000 to the creators and actually it's growing dramatically. So in a way we become like Airbnb for these cloud apps called actors right and actually this is like example of an actor it's a Google map scraper structure data from Google maps right and it is kind of things you would expect on the marketplace like reviews you know number of users rating and so on but the actors are actually represent a new way to ship software basically a piece of code that gets some input perform some action and generate some output right and actually this abstraction it's based on Docker.

So it's like standard sort of stack. But this abstraction allows us to generate for example SDKs for these actors, right? So it's it's super easy to integrate actors to your code to for example Python SDKs or JavaScript SDKs obviously like or like shell script because we have like command line interface as well. Open API we built actually a lot of integrations. So like it's super easy. We have like native integrations for example with Zepier or make and Slack and all kind of apps to make it really easy to call actors from your workflows basically and do the job like in the background for for your business. All right. And obviously recently MCP the model connect protocol from anthropic got a lot of traction. I haven't heard many people here talking about it. So now is the MCP time here. so it and it kind of like took the AI ecosystem by storm because like the adoption like grew dramatically and MCP is actually an like amazing amazing like u innovation because let's say we have like this store of 5,000 tools right to to that can be used by agents but you cannot expose 5,000 tools with all the input parameters like in open API function calling interface right basically the model wouldn't wouldn't be able to consume this, right?

Or like maybe like Gemini, but like it would like cost a fortune, right? But with MCP, MCP has like one super cool feature and it's like basically called like tool discovery, right? So you you expose a tool that kind of helps you find new tools and the client then registers those tools into the context and then can start using them, right? And actually MCP over the past like two months especially gained a lot of traction and actually became a de facto standard for inter agent communication right and I think it's mainly because the protocol is super simple so it's easy to to to build MCP server the servers are way way ahead of the clients I think the clients are only catching up now but as you can see like there's like 10 times more people are looking for MCP on Google than for open API right so I think it's clear open API is is not winning this battle and also while Google recently launched agent to agent interface or protocol and honestly I think MCP will prevail I think A2A will still work for some use cases but u MCP got rid of the critical velocity there are even like there are now like dozens of MCP server registries right there's even like from master AI our friends they built like MCP register registry just to find old registries of MCP server Right?

And you can see there is like 12,000 different MCP servers on the MCP market which is probably the largest. Right? So you think okay with MCP now the agents can interact with tools and that means like well now the agents can access the any of the any of those tools right well not really because still to access those tools you need to have API key for them right it doesn't work out of box so even if you use for example Zapier MCP server that provides like access to 5,000 apps you still need to connect those apps to your Zapier account right so basically everybody's talking about like how AI will you know take over the all the jobs and like you know take over big parts of economy but the agents are still not able to do like the very basic thing that people can it's like a find discover new services autonomously and purchase them right they just cannot do it right so I think if the agents are not able to do that like there will no there won't be this new agentic economy or like B2A where businesses sells to agents or agents sell to other agents unless we sort of solve this one problem.

Right? So what are the possible solutions of this? Right? So well in theory agents could subscribe to the services themselves, right? But for that they would need to have an email typically phone, credit card, some fake name and all that. And it's actually not very practical, right? like I think everybody who builds agents will attest to that that like this is just too risky and u and also you would need to breach the their terms of the services because obviously agents are not people right. So second solution is actually have some centralized identity and payments provider basically some central authority that where you will charge your account with with some money and then the agents actually can use that money to buy services right so there's a company like that called Skyfire I think this is great they are onto something but the problem is if you want to launch payment network like that it's very complicated it's like a chicken and egg problem right you need a lot of tools to support it and they will not support it until a lot of agents support it and they will not do it until the tools right so like launching the credit cards in in the ' 70s or PayPal in the '9s it cost a lot of effort and money right so I think this is this will take some time and effort right so I want to offer a third solution to this problem right and it's a what if there was like a centralized open marketplace of these tools and agents right and actually we already have marketplace like that with 5,000 tools and just recently over the last couple of days we added a capability that for example if there's an

external MCP server for some service that can be used by agents you can actually just wrap it as an actor and publish it in our store add monetization to it and that way you know you just the the agent just needs a API key to aify and basically the creator of the of the of the actor will put the credit card on the target service and make sure that it works right. So basically you just need one API key to access like 5,000 different tools and very soon maybe 10,000 of them right and I want to show quick demo we don't have much time so I will just do it u like from screenshots right so this is like our MCP client which is by the way another actor and I wanted to like ask it okay so give me a short summary of the AI rabbit hole conference speakers topics and so on so it will use coke tool or actor called rag web browser it's a web browser for you know like in GPT it will search for the query on Google and find the web page extract data from it and you can see notable speakers include my name is not there well obviously the models are still not great okay hallucinate quite a bit and okay another question is there an actor wrapping playright MCP server so it searches for the tools for the actors playright MCP and actually it found a tool because it's our new actor where we take playright MCP server which is like a product from Microsoft it's like browser automation library packages packages as an actor and then you can actually use it from from our parent MCP server from your agents right so yes there

is this tool there's some information about it okay so I ask it okay navigate to the conference website and take a screenshot all right so it it does that takes a screenshot screenshot of the conference that looks all right And I'm like okay. So we went even further. So we looked at one of our competitors is company called browser base. They also release MCP server like well everybody did right. So we were okay what if we can just wrap their MCP server as an actor publish on our marketplace and actually let anyone in aify ecosystem to use it. So we did just that. So now I'm actually adding the browser based MCP server to to the context of the client. It worked. Ra and now I ask ask it to do the same thing. Okay, also take a screenshot and and it did that right.

So basically now anybody can take an existing service or tool publish it on our marketplace and get you know out ofbox integrations to all the services and basically provide those services to the agents right so you don't need like 10,000 different API keys to all the services and you can just like your agents can now discover those through aify right so I went even a little further so I was like there's a form at the end of the web page fill it some fill it with some funny info and submit it. So, it scroll down and actually it filled it like the input search make. Well, it's not not too bad. All right. So, I'm running out of time, so I'll just wrap up quickly. so basically now anybody can publish these tools on our marketplace wrap them as actor add price tag to it and including our external agents and now they they will get exposure to like 700,000 monthly visitors that we get and more than 1 million revenue per months from this marketplace and you can tap into it very quickly.

Some open questions. Will the autonomous discovery provide real value? Everybody who builds AI agents knows that like it's easy to make a nice demo but it's quite hard to to make it work reliably right and and even with like fixed tools but if like the tools are you know discovered autonomously there is like even more volatility or like randomness right but I think over time the agents will go or the AI models will get so so so good that they will overcome that and they will become stable even with u autonomous discovery of tools. How can agents trust each other, right? Like I mean people and businesses when they in when they transact together, they need to trust each other, right? and we need to deploy this in a network of trust also to the agents to make this successful because like obviously there would be a lot of scam I can imagine like some fake scammer agents you know offering services for whatever and then stealing money and so on.

So obviously there is a lot of work you know to overcome this but this is my last question is really like what if this interaction of autonomous agents is actually what really achieves the AGI. So not like a one big data center like with like one huge model but really interaction of many different smaller agents very much like the intelligence emerges in biological systems right like people are built up from like simpler cells like companies are built up from people or like markets are built up from a lot of people right and this is a way how the intelligence emerges from the simpler components and maybe the AGI will emerge through interaction of agents but it will not happen if the agents cannot pay each other and really provide full autonomy because otherwise it would be the system would be fixed designed by people. So let's wrap up.

Thank you for your attention and we have a demo table in the lobby. So come and chat with us if you if you're interested.