MCP is the new front end for the agent internet

The fourth episode of Beyond the Bubble is a working session on what happens to the internet when the entry point stops being a search bar. Muzamil sits with Pierre-Louis Theron, co-founder and CEO of Alpic, a Paris-based company building an MCP-native cloud platform. Pierre-Louis has done internet infrastructure before. He co-founded Streamroot in 2013, helped keep French streaming platforms running during the 2018 FIFA World Cup, and ran Lumen’s global CDN as VP of Product after the acquisition. In July 2025 he reunited with his Streamroot co-founder Nikolay Rodionov to bet that the protocol layer for AI agents will be as foundational as HTTP was for the web.

The entry point to the internet is changing

The episode opens with Pierre-Louis stating the premise plainly. “The Google search bar has been the entry point of the internet for the past 20, 25 years,” he tells Muzamil, “but now the entry point is changing.” It is becoming AI chatbots and agents. The consequence is bigger than a UX shift. Every company that sells products, services, or even just presents itself online, he argues, will need to talk to AI in an AI-native way, and websites and APIs were not built for that conversation. Alpic exists to power that transition with tooling and infrastructure to deploy, secure, monitor and scale this new interface layer.

Muzamil pushes back early. ChatGPT already searches the web and reads pages. Why is that not enough? Pierre-Louis walks through how it actually works today: pre-training, then post-training and reinforcement learning, then a web index the model pings like a very fast human reading ten blue links. “It’s pretty dumb as well,” he says. Web pages were built for humans, full of UI, ads, pictures and buttons that exist for the wrong audience.

Booking a flight, and why links are a dead end

Pierre-Louis uses the canonical example: book me a flight to New York next week. Today the model returns links. You then go to the website. He argues the AI-native version of this looks completely different. The agent is connected directly to the back-end systems of Booking, Expedia, and the airline alliances. It returns options inside the chat. You book inside the chat. The redirect-to-website pattern, in his view, is “just something of the past.”

This is the moment he plants the flag on protocol. “This is exactly why I believe AI native protocol like MCP can really change the internet,” he says. Back-end systems speaking the same language as the model, instead of an LLM scraping a UI built for a human eye.

MCP is a front end for agents, not a replacement for APIs

Muzamil presses on the obvious question. APIs already exist. Coding agents are excellent at reading them. What is MCP actually for? Pierre-Louis acknowledges the confusion. When MCP first appeared, people declared APIs dead. Then MCP itself was declared dead. The maturation is now happening.

His framing is the cleanest part of the conversation. APIs present an exhaustive view of what a back end can do, “like a big menu.” MCP is something else. “At Alpic, we really see MCP as a new front end. It’s a front end for agents.” You design an MCP server with an intent, not for exhaustiveness. A single tool called “flight search” might wrap six APIs in the background and return a clean, agent-shaped response with pictures, prices, flights and airports. Without that, the agent has to discover and reason about a sprawling API surface every single time.

He is also clear about the failure modes. Context windows blow up when too many MCP servers are connected. He calls this an engineering problem, not a protocol problem, and points to “code mode,” implemented by Anthropic and Cloudflare, as one of the fixes already in flight: let the model call MCPs through code instead of stuffing the entire protocol exchange into context.

Websites are dead, and the kids will never visit one

Muzamil brings up a consumer agent he has been watching that manages family logistics, the kind of thing an executive used to have a human assistant for. His point lands sharply: if everyone gets that kind of assistant, the Google-to-website-to-purchase journey collapses. Trillions of dollars of value reshuffle.

Pierre-Louis agrees, and goes further. “I believe websites are dead.” He thinks the headless movement that started inside SaaS dashboards is about to eat the rest of the internet. His own two kids, he says, will likely never visit a website the way we do. He hedges only on timing and on whether it will be one super-agent or many. His instinct, drawn from how the industry consolidated around Windows and Apple, then Chrome and Firefox, then iOS and Android, is that we end up with a handful of agent platforms, probably ChatGPT, Claude, and one or two regional players.

He does not pretend MCP eats everything overnight. Your kid’s school will not ship an MCP server next quarter. Web search and agent-native interfaces will coexist for a long stretch before the agent-first internet fully takes over.

How a small business actually ships an MCP server

Muzamil asks the right next question: what does a non-technical operator actually do? Pierre-Louis grounds the answer in something familiar. MCP runs over HTTP. The Chrome-and-web-server mental model maps directly: ChatGPT is the new Chrome, an MCP server is the new web server.

He walks through the path. The most technical users can use open-source frameworks, including Alpic’s own Skybridge, point a coding agent at their API, and iterate the way they would on a front end. Then they need a place to host it. That can be existing infrastructure or a purpose-built platform like Alpic’s, which adds a playground, analytics, intent monitoring, and discoverability tooling. For smaller operators, off-the-shelf builders are emerging on top of those frameworks. He calls out the team at NoodleSeed as one of the players doing this work. “It brings back the power to the people with the ideas,” he says, “instead of only giving it to the people that have the technicality and the knowhow.”

Pre-Google, and the chicken-and-egg of discovery

This is the heart of the episode. Muzamil names the problem directly. Without discoverability, MCP is a novelty. Pizza shops will not change customer behaviour on principle.

Pierre-Louis accepts the framing. “We are right now at the equivalent of the internet pre Google,” he says. Before Google, you advertised your URL on billboards. That is roughly where MCP discovery sits today. He carves out one exception: B2B. SaaS users already live inside Claude, ChatGPT and Codex every day. For HR systems, CRMs, ERPs, shipping an MCP server is not a flex. As he puts it, “it’s a question of being alive a year or two from now.”

Then he describes the inflection he just witnessed in consumer. Claude has started doing what he calls dynamic discovery. He ran the test the afternoon of the recording. He asked Claude for pizza delivered to his office. Without touching web search, Claude returned three connector options including Uber Eats and Glovo. “That’s the first time we see that in the history of MCP,” he tells Muzamil. A leading model provider is prioritising AI-native answers over web search.

He backs the anecdote with data from an Alpic customer, Kiwi.com, the first online travel agency available in Claude. Eight months ago, putting an MCP on Claude looked like a strange bet because Claude was used mostly for coding. Once dynamic discovery turned on, Kiwi’s traffic through the connector, in his words, “literally exploded.”

Ranking inside chat is where Google was before PageRank

Muzamil pushes the harder version of the discovery question. The first 100 apps are easy: Spotify, Uber, the obvious brands. ChatGPT already has roughly 1,500 applications, up from about 100 in February. What happens when there are a million MCPs in the same category and the model can only surface three?

Pierre-Louis is honest that this is unsolved, and shows his work. By scraping Anthropic’s API, Alpic has spotted early signals: a rank field, a trending field that appears to track hits and new connections over a seven-day window, and inside ChatGPT, thumbs-up and thumbs-down widgets near each widget that could feed a feedback loop. Compared with the Google algorithm of 20 years ago, he is blunt: “that’s very, very basic.”

He flags two strategic implications. First, models know users far more intimately than Google ever did, which is a gold mine if handled with privacy intact. Second, because models will only surface a few options rather than ten pages of blue links, the prize for being early and well-described is real. Tool descriptions, he suggests, may become a new form of SEO.

WhatsApp, widgets, and the protocol bet

The conversation closes on a question Muzamil has clearly been sitting with. Agents are no longer confined to ChatGPT-style chat. They live in voice, in messaging, in vertical apps like the family assistant he opened with. How do you build one MCP that survives across all of them?

Pierre-Louis returns to first principles. The point of a protocol is that you do not rebuild for every surface. He expects protocol wars and consolidation, the same way browsers fought and settled. He singles out WhatsApp as a quiet contender for the agent layer, because friends, work and personal life increasingly happen there. Alpic, he reveals, recently demoed rendering an MCP app inside an iframe inside a WhatsApp chat. “A bit clunky and hacky, but we made it work.”

His advice to companies feeling overwhelmed by the surface area is two-part. Build one AI-native interface that any agent can reach. Then invest in being discoverable on whichever ecosystems matter, whether that turns out to be ChatGPT, Claude, an open agent world, or all of them. “People usually are lazy,” he tells Muzamil, “and they just end up in two or three places where they’ll feel comfortable.” The platforms will consolidate. The work is to be one of the few that get found when they do. As he closes, “the people that crack it early when it becomes big are going to win a lot.”