API Evangelist Conversation with Sameh Jarour on MCP, Actors, and Going Technical from a Marketing Background

My background is actually marketing and business — I worked across travel tech, cybersecurity, logistics, and e-commerce, and during the whole AI revolution I figured it was time to get more technical, because I could see AI was going to be the critical disruption of our time. I’d used Apify before for a podcast and loved their actors and tools, and when a product marketing opening came up I joined as the PMM for AI tools. In late 2025, when the MCP protocol came up, Apify was one of the real innovators in that area, so we got on the bandwagon right away with an MCP server. It’s grown well — on Claude Desktop we’re now number five in installs for the desktop extension. Everything runs in Docker, so we handle all the cloud infrastructure to run these actors and MCP servers in isolated sessions, and people just integrate with us.

One of the things we do well is breadth — there are several ways developers can integrate with us, whether from a CLI, from n8n, or from any AI company. We have a lot of actors, and every actor has its own input and output schemas, so it takes time to learn how to use each one. Because MCP reads the context of each actor, it’s easy for developers to test ten different actors quickly. If you’re doing Google Maps scraping there might be forty or fifty options; building a tool call per option would take forever, but with MCP you point your AI tool or IDE at all of them, test quickly, decide what you want, and then build something more deterministic around it. We were also one of the first to integrate dynamic tool discovery into our MCP server — you can even add an actor on the fly. It’s still ahead of where a lot of client compatibility is, but more clients are supporting it now.

The MCP server calls actors as tools, but there are also foundational tools in there — discover actor, find actor, fetch dataset. Those exist to help the client use the actor efficiently. We launched the MCP marketplace a few months ago, opening it up to our community — I think there are around 13,000 community developers now. They create actors and MCP servers and just actorize them. The reason Apify is so popular is the sheer number of integration points: you have one token, and now there are something like 30,000 tools behind it. It’s funny — our CEO, Jan Čurn, is super technical, and he’s the one who created the MCPC protocol, a universal CLI client for MCP that tries to solve the compatibility issues different clients have, like a Swiss-army knife you plug into any MCP server to discover its capabilities before you test. He vibe-coded it over the Christmas break. That says a lot about the company.

If you compare it to two years ago, the addressable market for developers has just exploded. Before, maybe ten in a thousand people were developers; now you can classify a lot more people that way, because the vibe-coding revolution plus the advancement in the models has made it much easier for people like me. And it’s not only about production — building applications and agents — it’s also about learning. I’m learning every day from my interactions with Claude. It’s been transformational for a lot of people in marketing and business-oriented functions at Apify, so I’m not the exception. I came from being a marketing guy, and now I’m vibing on Claude Code for most of my workday, with a bunch of Claude agents running on my Mac mini. The exceptions now are the people who aren’t building applications or agents.

That’s exactly it. Usually people have their code siloed somewhere, but once they publish it into Apify it becomes an abstraction layer that gives them access to millions of users and to integrators. That’s a key part: it’s super easy to integrate once you publish on Apify, so it gives you market reach and visibility. And I think AI and agents are forcing a lot of people to make their services more accessible whether they like it or not. With Apify you don’t necessarily have to build the API yourself — you can publish almost anything, and we’ll create an actor abstraction around it so it can actually be accessed and put to work. Combined with a tool like Claude Code, that’s a real shift in how people get their capabilities out into the world.

Luckily I have clients whose leadership is pushing AI very aggressively. There are real concerns around data security, especially in the GCC, so they tend to want to do things internally from a data-residency point of view — a big topic right now is whether you’re sending training data to these models. But they still need external knowledge, because building multi-agent systems inside enterprises is pretty new, so they rely on consultants like me to help with model choice, RAG pipelines, best practices on which tools and APIs to use, and how to conduct evals. At the same time we’re building capacity inside their IT and technology teams, because they see this as existential — they need to build it out and scale it up in the next few months if they want to stay competitive.

Yes, exactly — that’s a forward-deployment model. The big thing I’m building right now is, for example, voice AI agents for an e-commerce and freight client, taking their call center and transitioning it into voice. But more broadly, every custom application we build, we make sure every service is an API call and that we have agents that can call that service and integrate it with whatever channels they need. Custom applications are the big thing now, which is why a lot of SaaS vendors are losing market share — they had to build platforms general enough for everyone, full of features you never required, but never specific enough. With vibe coding and agentic models you can build your own custom platforms, expose every feature as an API documented with markdown files, share those across the agents, and let users themselves bring in or turn off the features they want.

I don’t think the displacement is as fast as people expected — the models are great but they still hallucinate, and there’s still a lot of resistance from human users. I thought by now most white-collar jobs would just go away, and that hasn’t been the case. But people like me, who have a commercial background in sales and marketing and then became technical, are probably the ones who are going to mint in this environment, because we can lean on AI agents to help with the code. The biggest shift I’ve seen, at least in product marketing, is going from features-centric to outcomes-centric. We used to talk about features, target markets, churn, and incremental income. Now that everyone can build their own software, we don’t really talk about features anymore — we talk about the outcome: how is this growing my top line or decreasing my bottom line?