What Happened When I Put the Promise of MCP to the Test of Real Life?

The Truth Behind The Hype 

 
Everyone is talking about Model Context Protocol (MCP).  

The promise – let models talk directly to systems instead of building endless glue code. I wanted to see how it can solve several existing challenges, both internally and for our customers. 

So, I ran two real-world projects: 

• Text‑to‑SQL on a PostgreSQL database (5 schemas, 20 tables) for a global financial services provider. 

• Internal Zendesk analytics for Marketing and Professional Services at 2bcloud. 

Both solutions moved beyond theory; they’re running in production today.

Chapter One – Making SQL Queries Human 

The first challenge came when a customer asked 2bcloud:  

“Can our analysts query financial data in plain English?” 

This wasn’t a toy dataset. The client’s business spans Equity, Fixed Income, FX, and specialized strategies across 90+ trading venues worldwide.  

Underneath: Amazon RDS PostgreSQL database with five schemas and 20 tables.  

AWS’s Way vs. Reality 

I started with AWS’s recommended Text‑to‑SQL architecture with RAG technique. 

It looked strong on paper: Amazon Bedrock models, error correction, multi‑source connectors.  

But in practice, it required OpenSearch, Glue Data Catalog, and Athena with heavy error handling. Solid, but way too much machinery for this client’s needs. 

The Turning Point: MCP Breakthrough 

That’s when I revisited Anthropic’s MCP. The promise of models speaking directly to systems made sense. The real turning point was when AWS announced MCP server support for RDS PostgreSQL. Exactly what I needed. 

Getting Hands Dirty: Claude & Flowise 

I spun up a sandbox with AWS RDS PostgreSQL and test data.  

First, I tried Claude Desktop MCP Client with Sonnet 3.7. Integration was smooth and gave me a feel for MCP, but it was not deployable to customers since the MCP Client deployment required a customer environment. 

Next stop: orchestration tools. Langflow was interesting, but Flowise was more intuitive. I wired up Claude Sonnet through AWS Bedrock into Flowise; the response format wasn’t Flowise‑friendly. Sometime later, I realized that it needed to be in OpenAI format. The fix: use LiteLLM as a bridge. Problem solved. 

Real Queries, Real Answers 

Did it work?  

Better than expected.  

Despite a heavy schema, the model handled queries in 2–5 minutes: 

• “What is the average EBITDA multiple of UK M&A deals in the REIT sector and what were the deals?” 

• “What was the total enterprise value of German deals in 2023?” 

• “What country had the most M&A deals by number in 2024?” 

• “How many European deals were subject to US antitrust in 2024?” 

• “How many new clients were added to the European distribution list this year?” 

• “What is the average closing time of European deals launched in 2024?” 

• “What was the most common sector of European deals this year?” 

• “What is the average premium of Italian bank sector deals?” 

For production, we deployed on AWS Fargate, scalable and reliable. 

Chapter Two – Turning MCP on Ourselves 

The next use case was internal.

• Marketing wanted to mine Zendesk tickets for Azure and AWS service trends.

• Professional Services wanted service efficiency and quality insights. 

Their existing method – exporting tickets to JSON and running a RAG pipeline – was brittle and shallow. It couldn’t provide dynamic analysis. 

Choosing the Right MCP Server 

I tested community MCP servers and landed on zd‑mcp‑server.  

Why? It was built for internal teams and had the one feature we needed most:  
“Search tickets with query syntax.”

For the model, I used OpenAI o3. 

Asking Zendesk the Right Questions 

This setup lets us run: 

• “Number of tickets per person for the last month” 

• “Number of tickets closed last month, by person” 

• “Are there any tickets awaiting feedback” 

• “Assess customer sentiment for Customer X” 

• “Identify trends from tickets opened in the last 3 months” 

The answers were solid. Once again, we deployed through Flowise to make it production‑ready. 

Final Thought: From Buzz to Production 

Two very different problems – financial data analysis for a global client, and Zendesk insights for our internal use – came back to the same lesson:  

MCP stripped away the overhead and let models talk to systems directly. 

The hype around MCP is everywhere.  

My test was simple: does it work outside of theory?  

The answer was: yes. 

— 

Mark Valman 
Cloud Solution Architect @ 2bcloud 
 
Want to learn more about my MCP experiments?  
Contact me [email protected]