MCP Server Development Case Study

Client & Overview

The Client & the Problem

The client is a small, Italy-based travel company that aggregates and sells train, bus, and air tickets. Their booking engine sources pricing from multiple transport suppliers for routes across Italy, and they wanted that process to be simple and transparent enough for their users to see live pricing directly through AI agents. Their product had no way to do that: a user asking Claude or another AI agent about a transfer fare got nothing, because pricing lived behind the client's own backend, not behind any interface an AI tool could call. TechAvidus was brought in to build a standalone Model Context Protocol (MCP) server sitting in front of that backend.

Obstacles & Solutions

Business & Technical Challenges

While developing the system, we had to address the following core challenges:

Business Challenge: Every time an end user asked about transportation pricing, the client had to manually check six different suppliers one by one in their software. The client wanted a single response returning pricing faster than checking each provider individually.

Normalize Suppliers: Normalize six suppliers' structurally different pricing feeds into one consistent, AI-callable response shape.

Dual Runtime Modes: Support two runtime modes—local stdio for development and public streamable-HTTP for production—from a single codebase without duplicating tool logic.

LLM Optimization & Security: Compress verbose backend JSON into concise, deterministic text an LLM could reason over, and scope a publicly exposed endpoint securely without widening backend exposure.

GET A FREE CONSULTATION

Solution Architecture

A Standalone FastMCP Server, Built for AI Agents

TechAvidus built a standalone FastMCP server, written in Python, that exposes four travel-pricing tools to AI agents like Claude. Rather than rebuilding the client's supplier aggregation and pricing logic, the server connects to three of the client's existing backend REST endpoints and supports both local stdio and hosted streamable-HTTP transport from one codebase.

The project was delivered by one MCP developer over roughly two months and concluded at handover. The existing frontend, backend data pipeline, and supplier-level pricing normalization all remained outside the engagement.

Security & Architecture

Security, Protocol, & Execution Pipeline

The client had no documentation of their existing product, so the engagement started with direct conversations and hands-on examination of their existing system. Three foundational decisions shaped the build:

Security Scoping Before Public Exposure

The server ships with TransportSecuritySettings, restricting allowed hosts and origins to prevent the endpoint from acting as an open proxy in front of the client's backend.

Keeping the MCP Layer Thin

Price aggregation, normalization, and caching stay in the client's backend. The MCP server handles tool exposure, parameter validation, and response formatting only.

A Layered Execution Pipeline

Each tool call moves through discovery, parameter validation, backend call, and LLM-formatted response sequentially to support dual transports cleanly.

Core Features & Architecture

An AI agent calls one of four MCP tools; the server validates the request, calls the client's backend over async httpx, and formats the response for an LLM to read. It holds no pricing logic, no caching, and no data of its own.

Route Discovery

Transport Search

Exposes search_italy_transport, list_italy_routes, and get_route_info tools, formatting grouped, capped-supplier-list responses for natural-language queries.

Intelligence

Price Calendar

Exposes get_cheapest_travel_day to return day-by-day cheapest/average price breakdowns and optimal travel day flags, with graceful error fallbacks.

Security & Transport

Dual Transport Layer

Runs identical tool behavior over local stdio and public streamable-HTTP with environment-based API key authentication and host/origin restrictions.

System Overview

Technology Stack

Languages & Frameworks

Python, FastMCP

Client Libraries

httpx

Async HTTP client

Protocols / Transports

MCP & Stdio

FastMCP streamable-HTTP transport

AI Client Compatibility

Claude Desktop, Cursor, VS Code

Security

TransportSecuritySettings

Host/origin restriction & API key

Get Started

Need an MCP Server Built for Your Platform?

We help businesses expose data and workflows to AI agents through custom, secure, and production-ready Model Context Protocol servers. Whether it's live search, pricing models, or internal tools, we build clean integration layers that AI models can reason over effortlessly.

GET A FREE CONSULTATION

What Our Clients Are Saying

I worked with TechAvidus for over 2 years. The team is incredible. They are very responsive, responsible, dedicated, and capable. I strongly recommend the team for any project, however complex.

Alexander Z Stanley

Business Development Manager

Keshu, Bhavesh and their team are outstanding. Not only were they able to build a very sophisticated and customized portal for our thousands of customers, but they did so with skill, efficiency, professionalism, and offered wonderful suggestions and recommendations along they way. Even more impressive, they were able to do all of this in spite of the convoluted direction we provided them with.

We've since expanded our relationship with them to cover many additional web development jobs. Choosing to work with TechAvidus was one of the best decisions we've made in recent years. We recommend them unconditionally.

Rob Price

President, Gatekeeper Press

"The vendor did an outstanding job. They paid attention to every detail of project scope and made prompt corrections when needed. I would definitely work with them again."

Jason Lorenti

Digital Development Manager at Evolved Office LLC

Love to work with the team with a fair pricing and very valueable developers for komplex usecase. Actually we worked on a really big module based application and every work is awesome done.

Matthias Grdan

CEO | Digital Transformation

They were very responsive and the quality of the work was good. We experienced a few challenges with the back and forth during deliveries and some unexpected delays but they were always receptive to feedback and responded well to our concerns. I'd happily recommend them, especially Megha as the Project Manager and Savan as the Developer.

Marcos Queiroz

Director at Rapido

View All Global Reviews

We're the cookies !

Decide how Cookies uses your data

Toggle all

A Production MCP Server Unifying Supplier's Travel Pricing for AI Agents

Our Clients