The rise of the agentic economy on the shoulders of MCP — Jan Curn, Apify

0:01[Music]

0:14So let me start with a question. How

0:16does intelligence emerges in biological

0:19systems? Right? Well, it's through

0:22neurons, right? When neurons are born,

0:24they are just like individual cells. But

0:27like over time, they grow their axons

0:29and dendrites and establish connections

0:31with other cells or other neurons and

0:34actually learn how to communicate in

0:36order to pursue their own interest

0:38basically like to get nutrients and so

0:40on. And over time they learn how to how

0:42to communicate with each other and with

0:43other cells to get nutrients and

0:45basically thrive, right? And this

0:47collective behavior if you like zoom out

0:49and look at like really large number of

0:51them uh is something we call

0:53intelligence right so it's like emerging

0:55behavior of smaller individual units

0:57that pursue their own interest.

1:00So how does intelligence emerge emerge

1:03in the markets right people always talk

1:05about markets like well market thinks

1:07that market uh reacted to this and so on

1:11and in some way uh markets are more

1:14intelligent uh than like individual like

1:16participants of the market right and

1:18it's there are mutual uh interaction of

1:22these individual members of the market

1:24basically who pursue their own interest

1:26and communicate and establish new

1:28interactions with others

1:30uh where some some sort of like

1:31collective intelligence which is like

1:33bigger than the sum of different parts

1:35emerges right so how does intelligence

1:37emerges uh emerge in companies well this

1:41one is provocative through slack right

1:44where people interact and pursue their

1:46own interest uh in the company and over

1:48like altogether the company well

1:51sometimes becomes more intelligent than

1:53the individual employees of the company

1:55and uh so this leads to my final

1:58question so how or how will the general

2:00intelligence emerge in computing systems

2:02right and there is a lot of talk about

2:04AGI and like you know like ever larger

2:08models uh exhibiting like super

2:09intelligent behavior but in my opinion

2:12the like general intelligence will

2:14actually emerge through interaction of

2:16multiple entities can call them agents

2:19basically like multiple models uh

2:21pursuing their own goals interacting

2:23with each other and uh altogether

2:25exhibiting something which we can call

2:28general intelligence

2:30and thanks to uh MCP we finally have

2:34this uh missing part that allows the the

2:37agents to communicate with with each

2:38other and really like create a fabric or

2:40agentic mesh where they can talk

2:42together. So uh hello everyone my name

2:45is Yan Churn. I'm the founder of API and

2:48uh I'm going to talk about the race of

2:50the agentic economy on the shoulders of

2:52MCP basically economy where agents uh

2:56can you know find counterparts uh to

2:59interact with and purchase services from

3:02other from businesses or tools or other

3:04agents right so like B2A and B2B uh

3:07sorry and A2A

3:09all right so before I start um let me

3:13just introduce quickly API aify is

3:16is a marketplace of 5,000 tools called

3:18actors and uh historically we come from

3:21the web scraping industry right so most

3:23of these actors are data extraction

3:24tools that allow you you know to get

3:26data from social media from search

3:28engines uh data for AI for building rack

3:31pipelines you know uh data from web uh

3:34for lead generation and so on but also

3:36there are other tools like data

3:38processing tools and so on so altogether

3:40there's about 5,000 of them and some of

3:42them are built by API some are built by

3:45our community of creators uh who

3:47actually make money on it, right? So

3:48it's like a marketplace of software

3:51creators if you will right so actors are

3:55self-contained piece of pieces of

3:56software based on docker with well

3:58definfined input and output right and

4:00basically they represent a new way how

4:02to ship software and publish it you know

4:03and uh and integrate to to you know

4:06other systems right so for example

4:08Google map scraper it's a quite popular

4:10actor uh from our store uh it can

4:12extract data from Google maps right uh

4:14more data than than the Google places

4:16API provides right

4:18Well, there is like creator of the

4:19actor, description, you know, different

4:22stats and so on. Something you would

4:23expect from normal marketplace. And

4:25actually,

4:28thanks to the way actors are built, it's

4:30actually super easy to integrate actors

4:32from other systems, right? Right. So for

4:33example we have SDKs for TypeScript for

4:36Python uh for open API uh for C common

4:40uh for CLI like we can call them from

4:42terminal and it's only because they are

4:44well definfined units of software with

4:46input and output right. Uh also we have

4:49integrations with uh workflow automation

4:52tools like mag zapier you know clay and

4:54many others. So to make it really easy

4:56to call actors from these systems,

4:59right? But obviously now uh we also have

5:02MCP integration which makes it possible

5:04to call actors from AI agents or you

5:07know AI workflows. And the way it works

5:10actually is uh the agent just needs a

5:14API key or you know o workflow on an an

5:18account on aifi and then through our MCP

5:21server basically it can interact or call

5:25any of those 5,000 actors on our

5:27marketplace right and actually this only

5:29became possible thanks to uh I would say

5:32the killer feature of MCP which is the

5:34tool discovery right actually um not

5:37many clients didn't support this yet. Uh

5:40but uh just just today I saw that V VS

5:42code added support for it. Uh and

5:45actually just like two days ago cloud

5:47for desktop added support for tool

5:48discovery. And basically how it works is

5:53that um the client connects to the MCP

5:56server and dynamically discovers tools

5:58to use and to interact with based on the

6:01based on the the workflow, right? And

6:03let's say we have like 5,000 tools on

6:06our our store and there is simply no way

6:08we could publish all these tools through

6:09open API because you know the context

6:11would be just too large and like the

6:13more tools you have the you know riskier

6:16the result is right so we really want

6:18like provide the tools only like uh as

6:20needed and that is only possible through

6:23tool discovery which I think is really

6:25the main thing that will actually make

6:28MCP really uh

6:31the huge differentiator from from open

6:33API for example, right? So MCP actually

6:37quickly became a standard for agentic

6:38interaction. This is Google trends data

6:41showing that MCP is is basically

6:43dominating the space compared to open

6:44API or A2A from from Google, right? And

6:48actually I think MCP already became a

6:50standard for agenting interaction.

6:53And it became so popular that currently

6:55there are like you know many different

6:58like uh registries of MCP servers that

7:01even guys from master our friends

7:03created like registry of MCP server

7:05registries right just to make the sense

7:07of it right and obviously Antropic is

7:09also working on their own uh registry

7:12and um I think Google's A2A they have

7:15like a DNSbased protocol with like

7:16well-known agents JSON way to you know

7:20publish the the services on through DNS.

7:24So basically there is like you know so

7:25many different servers you can now use

7:27from the agents right so does it mean

7:32with like so many tools now support MCP

7:34so does it mean like the agents can

7:37discover and access any of them on their

7:39own right

7:41well not really because

7:44to use those services your agents still

7:46need uh to have like API tokens to those

7:48services right so even let's say if you

7:51use Zapier MCMCP that provides access to

7:54like 5,000 apps they have in their

7:55marketplace. You still need to connect

7:58those individual apps to your services,

8:00right? You know, like GitHub or Slack or

8:02you know whatever. So, Zapier on the on

8:05its own is not able to provide access to

8:07the third party services. You still need

8:09to as a user to facilitate that. So that

8:12actually means that uh the agents are

8:15not able to like find counterparts uh or

8:19like uh other agents or other tools to

8:22interact with on their own. They are

8:24still depending on the human uh

8:26developer who actually build the system,

8:27right? Who kind of like give those those

8:28those agents access to different tools,

8:31right? And if those agents are, you

8:34know, to replace all the people and all

8:36the jobs, right? they need to be able to

8:39uh find services to interact with. They

8:42can't just like you know do do that like

8:43it's like basic basic thing that like uh

8:46anyone of us can do right like to find

8:49service and purchase it right. So I

8:52argue that like unless the the the

8:54agents are able to do that uh we will

8:57not be able to reach you know some

9:00higher level of of intelligence of these

9:01agentic systems and behaviors basically

9:04uh if the agents cannot purchase

9:06services right so how can we solve this

9:09problem right so first like sort of like

9:11n approach would be let the agents

9:13subscribes themselves to the target

9:16services right so basically in a way

9:18like agents could have like email maybe

9:20credit part they could like fill you

9:22know the subscription flow maybe solve

9:24the capture you know create an account

9:25and so on but you see it's it's not very

9:27practical right I mean it's you know

9:31well they might need to also have to

9:32phone number and so on and quite often

9:34the services actually need to have like

9:36real person behind the account right so

9:38basically this this wouldn't really work

9:39right

9:41uh so second solution uh is central

9:45identity and payments provider there are

9:46like couple of companies pursuing now

9:48that like there would be like a central

9:49authority where you and charge money and

9:51then the agents can use that you know to

9:53buy services and and provide them with

9:55their identity right for example

9:57vertifier coinbase is now pushing their

9:59X42 standard uh I think Stripe is

10:02working on this and Mastercard and Visa

10:042 right so I think this this is going to

10:06happen eventually but running launching

10:09new payment system it's extremely

10:11complicated right because you're facing

10:12like this chicken and egg problem of

10:14marketplaces right I think PayPal had to

10:17uh pay like $100 million per month just

10:19to buy the market uh and like launching

10:21credit cards in the in the 70s was like

10:24incredible challenge basically because

10:25nobody was accepting those cards so why

10:26would people use them and so on right so

10:28I think this will happen but it will be

10:30long process basically to establish this

10:32right so let me offer the third approach

10:35and it's like through a centralized

10:37marketplace of MTP services like ampify

10:40store basically where you just need one

10:42API token or one authentication one

10:44account to to get access to all the

10:46other services and basically it works

10:48the way that the developers who publish

10:51these tools, these actors actually they

10:53provide their credit card and their

10:55account to the third party service and

10:56basically publish it, add monetization

10:59to it like how much does it cost to call

11:01the service and then they are basically

11:03the owner of the service and now they

11:05publish it on our marketplace and

11:07suddenly it becomes available to the

11:08whole ecosystem of tools and this way

11:10actually we can scale it rapidly and

11:12actually even without the target

11:14services knowing right so basically this

11:16way the actor can run the code itself or

11:19wrap an external API or just publish an

11:21external MCP server because the MCP

11:24servers they can be actually nested you

11:25can have like one parent server that

11:27provides actions or tools of the like

11:30nested MCP servers right so that's

11:32another like cool feature of FMCP you

11:34can really build this sort of ecosystem

11:36you know if you can facilitate the

11:37payments and monetization right so

11:41actors charges the user and then its

11:43developer gets the money and pays for

11:45the external service and anyone can

11:47publish such connector even without the

11:49target service knowing right. So time

11:52for demo it's not live demo because the

11:54internet uh super flaky here. So what

11:57you can see here is clot for desktop uh

12:00that has access to API MCP server uh

12:03there is like 18 tools available now and

12:05I'm asking like what is the venue of AI

12:07engineer world's fair in San Francisco

12:09it possible use actors so it you can see

12:12it searches the actors

12:14uh for a tool that can answer this

12:17question it will find a tool or actor

12:19called rag web browser and so it's

12:22called it's a it's like a Google search

12:24with you know fetch data so basically It

12:26it asks the query like uh what is the

12:28venue and so on and then it parses the

12:30the resulting page. So we can see it

12:32found like uh SF Marriott Mary's uh that

12:36seems all correct, right? So now let's

12:39use an actor for scraping Twitter.

12:43So uh this actor is not available in the

12:46context. So so the agent doesn't know

12:47how to use it. So it will it searches

12:49actors on our store and finds an actor

12:51that can scrape Twitter, right? So it it

12:55calls it calls add actor which is like a

12:57tool that adds new tool to the context.

13:00Uh actually cloud is very verbose

13:02describing a lot of things about it. And

13:04actually there is like small bug still

13:05in cloud escope that you need to like

13:07disable and enable a tool so that the

13:09the tool list refreshes and then the

13:11tools become available. I'm sure it's

13:13going to be fixed in the next release.

13:15And now let's use that actor to uh get

13:18the last tweet of AI engineer

13:19conference. All right. So it calls the

13:22actor on aify. Uh it knows the the

13:26Twitter handle probably from from the

13:27from the website. And now you can see

13:30that uh it found the result and the last

13:33tweet from this morning was uh something

13:36about workshops. That seems about right.

13:38So now what? So we we have we have seen

13:42how we can use existing tools in in in

13:44our store. But like let's say uh uh one

13:49of our competitors uh company called

13:51browser base. Hey Paul if you're here uh

13:54they certainly haven't published you

13:56know uh an actor in our store but we

13:58did. So we created an account browser

14:00base, added our API token there and

14:02published like basically their MCP

14:05server on our store without actually

14:08them even knowing. And now anybody can

14:10actually use browserbased MCP through

14:12API's ecosystem, right? Even without

14:16them having to do anything or knowing

14:17about it, right? So now let's use

14:19browser base to fill in the email

14:21subscription form on the AI engineer

14:24website.

14:26fill email uh yanappify.com

14:28and now let's see what happens right and

14:30actually we'll see that uh that the that

14:33the agent will actually call

14:35browserbased MCP through you know an

14:38actor publish published you know by us

14:40on our oni store and perform the actions

14:42on the web right and actually this way

14:45we can easily like uh bring a lot of lot

14:47of existing MCP servers to our store and

14:50you know expand the ecosystem rapidly

14:52without you know having to ask for for

14:55you cooperation of the third parties.

14:57Right? So that's actually what we're

14:59doing now. Uh we want to scale this

15:01marketplace rapidly and now okay so now

15:05it's evaluating you know the screenshots

15:07looking for the field and so on you know

15:09and eventually it will manage to uh fill

15:12the form and and basically succeed in

15:15the task. Right? I can maybe skip this

15:18uh to save time. It takes some some some

15:21time to basically

15:23uh for the agent to to find the form and

15:26so on. But uh yeah,

15:29it succeeded. It completed the email sub

15:31subscription. And

15:34this way you basically see that uh you

15:37can plug

15:40our our ecosystem of actors into into uh

15:43any AI agents that actually support tool

15:45discovery. Right. All right. And

15:52so this means now anyone can publish

15:56tools or you know agents on aify store

15:59and monetize them and immediately get

16:01access you know to all the AI clients

16:05that already like integrated appy and

16:07all the ecosystem of tools right and

16:09actually people can make make money on

16:11it like just last month we paid more

16:13than4 million dollars to our creators

16:15and actually the the this number is is

16:17growing. rapidly. You know, overall the

16:20actors generate more than one half

16:21million dollars per month. Now, uh we

16:23have like 1 million monthly visitors to

16:25the whole ecosystem. And now we're

16:26really in the process of like scaling

16:28this ecosystem. So, um if you're looking

16:31for ways to monetize your tools or

16:33agents, you know, just um talk to us and

16:36publish or notify store and get access

16:39to this ecosystem of developers and this

16:42visibility.

16:44And there are some open questions

16:46obviously uh that remain. So will this

16:50autonomous through discovery provide

16:52real value? I mean like everybody who

16:54builds agentic systems knows that you

16:57know like making sure that the system

16:59works as expected is tricky right even

17:02if it's fixed. So if we add this like

17:05you know variables that like well uh if

17:08the agents can discover new tools uh

17:10will it actually work? Well, currently

17:12it might it might be a bit flaky, right?

17:14I think we're still still we we're still

17:16fairly early, but as the models get

17:18better,

17:20I think uh even with the discovery,

17:24suddenly the the the agents will be will

17:26be able to provide you know valuable and

17:29reliable result basically right so this

17:32is this remains to be seen but I'm

17:34optimistic that like as the LMS will get

17:36better we'll actually get there that the

17:38two discovery will actually provide real

17:40value. Well, there's a big question of

17:42like how can agents trust tools or other

17:44tools? Oh, so sorry. Or each other,

17:46right? We know it like you only interact

17:48with people you trust. So, how can

17:49agents do that? We'll see. And can

17:52autonomous agent interaction enable AGI?

17:54Well, we'll see. Thank you very much for

17:57your attention and uh feel free to try

17:59it. MCP.I.com.

18:02[Music]