AI MCP Events Paris Agents

MCP Connect Day: Notes from Paris

Last week I attended MCP Connect Day in Paris, organized by the team at Alpic. It was a full day of talks from engineers at OpenAI, Hugging Face, Mistral, GitHub, Postman and others, all centered around the Model Context Protocol and what’s being built on top of it.

Here’s what I walked away thinking about.

What the day covered

The talks fell into a few natural clusters that I found genuinely useful:

  • The protocol itself: where MCP is today, where it’s heading, and what’s still unresolved
  • Building MCP servers at scale: lessons from teams running them in production for millions of users
  • ChatGPT Apps and generative UI: how the UX layer on top of MCP is evolving
  • Discoverability: the underappreciated problem of making your MCP server findable to LLMs

MCP is maturing, but the hard problems are protocol-level

Shaun Smith from Hugging Face gave a talk on transport evolution that stuck with me. Hugging Face hosts thousands of MCP servers and has real traffic data, including a humongous spike when Claude AI integrated remote MCP, followed by a drop when clients were behaving inefficiently.

The upcoming changes to the protocol are significant: moving toward stateless requests (every message makes sense in isolation), elevating sessions as a first-class concept via a cookie-style mechanism, and fixing the scaling problems around server-to-client communication. These aren’t theoretical. They’re responses to real production pain.

The takeaway: if you’re planning to run MCP servers at any meaningful scale, the architecture decisions you make today will matter. Sticky sessions are a crutch. The protocol is moving away from them.

MCP is not just an API wrapper

Kevin Platel from Leboncoin made this point clearly. They serve 30 million users and went through the exercise of building a ChatGPT app. The insight that resonated most: when you build an MCP server, you’re not translating your API for machines. You’re building for a new kind of user: an LLM that reasons about intent, not endpoints.

Their conclusion was that MCP is a new distribution channel, similar to the mobile era. The companies that succeed won’t be the ones who port their REST API over. They’ll be the ones who rethink the product flow entirely.

The unsolved problems they named: discoverability (how does a user know your MCP app exists when there’s no search bar?) and trust in transactions (how do you build confidence around money and goods without the familiar visual cues?).

Generative UI is an open question

Ruben Casas from Postman gave a talk I found thought-provoking on the spectrum of what “generative UI” actually means. The short version: most of what’s being called generative UI today is still static components with dynamic data. The genuinely generative version (where the LLM writes the UI at runtime) is possible and he demoed it, but it requires a sandbox.

The framing I found useful: ask yourself what the LLM is rendering. Parameters and data? Static components. JSON descriptors? Declarative UI. Actual HTML/CSS/JS at runtime? That’s where it gets interesting, and also where it gets risky without proper sandboxing. MCP apps are an interesting fit here precisely because the iframe model gives you that sandbox out of the box.

Discoverability is the problem nobody has solved

A few talks converged on this. One was bluntly titled “If the LLM Can’t Find You, You Don’t Exist.” Sam Morrow from GitHub talked about how they’ve gone from 100+ tools to ~40 defaults, not because fewer tools are better in theory, but because users don’t configure things. They use defaults and complain.

The pattern across talks: most MCP configuration decisions that should be solved at the protocol level are instead being solved by individual server maintainers papering over gaps. Tool grouping specs keep getting rejected. Dynamic discovery is client-specific. It’s a rough moment in the ecosystem’s maturity, which is also exactly the right time to be paying attention.

Where I landed

A few things I’m taking back with me.

MCP as a concept is sound. The problem isn’t the protocol idea. The ecosystem is still early and the tooling, clients, and specs are all moving at different speeds.

The intent-oriented design point from Leboncoin is important for how we think about our own work. We’re not exposing an API. We’re designing an experience for an agent that reasons differently than a human user does.

And the community is genuinely good. There’s a real group of practitioners working through the same problems and sharing what they’re learning openly. Worth staying close to.

Attended: MCP Connect Day, Paris, 8 February 2026