🎤 Speaking at GDG Firenze - Model Context Protocol

Posted Mar 5, 2026 Updated Mar 9, 2026

By Enrico Nello

3 min read

what an incredible evening!! on march 5th i had the pleasure of speaking at the Google Developer Group Firenze meetup, talking about the Model Context Protocol (MCP) in front of a packed room at 42 Firenze.

the energy in the room, the questions, the conversations that kept going during the networking session after — it was a fantastic reminder of how vibrant and passionate this community is.

🎯 The Talk

“Model Context Protocol: l’importanza della standardizzazione nell’era degli agenti AI”

the core theme of the session was one i feel strongly about: standardization matters. in an AI ecosystem as fragmented as today’s, without shared protocols the risk is getting trapped in endless custom integrations — closed ecosystems that slow down innovation instead of accelerating it.

this is exactly why MCP is establishing itself as the de facto standard: a unified protocol that allows AI models and agents to access tools and data in a uniform, secure and interoperable way.

🗂️ What We Covered

the session was structured to move progressively from theory to practice:

The Current AI Ecosystem

a brief overview of how we got here — the fragmentation problem and the path that led to MCP: what it is, why it was created by Anthropic, and how its architecture works.

Interacting With MCP Servers

starting from the simplest way to interact with MCP servers: using them directly inside chatbots and IDEs. understanding the basics before diving deeper.

Building Custom MCP Servers

from consuming servers to building them — how to define and create custom MCP servers designed for specific needs. different deployment strategies were explored:

local execution for development and testing
cloud deployment for production use
third-party MCP servers from the growing ecosystem

Multi-Agent Orchestration

the final and most advanced part: building a multi-agent system that combines multiple MCP-based tools, showing how agents can collaborate intelligently within complex workflows.

the demo first showed a hybrid agent connecting to 4 MCP sources simultaneously — a local recipe server, a remote weather server, a Docker MCP Gateway for YouTube transcription, and a Hugging Face Space for image generation. this immediately exposed a real problem: tool overload. too many tools confuse the agent, degrade performance, and make routing unreliable.

the solution presented was a triage-based swarm architecture (source), where a TriageAgent acts as the orchestrator and hands off each request to the right specialized subagent:

TriageAgent (orchestrator)
    ├── RecipeAgent    → local MCP server (italian recipes)
    ├── WeatherAgent   → FastMCP cloud server (weather)
    ├── VideoAgent     → Docker MCP Gateway (YouTube transcription)
    └── ImageAgent     → Hugging Face Space as MCP tool (image generation)

each subagent only has access to its own tools — clean separation of concerns, better performance, and easy to extend. the whole system was built with LlamaIndex for agent orchestration, FastMCP for the custom server, and a Gradio UI for the live demo.

the multi-agent swarm architecture presented during the demo — full code available on GitHub

💡 Key Takeaways

standardization is not optional — without it, AI ecosystems fragment into silos
MCP is the universal adapter for connecting LLMs to tools, databases, APIs and services
once you understand the protocol, building and deploying custom servers is surprisingly straightforward
multi-agent systems built on MCP can tackle complex, real-world workflows with clean separation of concerns
the community interest was huge — from developers already experimenting with MCP to people discovering it for the first time