My POC with MCP: Connecting Claude with a Test API

2 minute read

POC with MCP: Connecting Claude with a Test API

How it all started

It was 4 AM and for some reason I couldn’t sleep. I was scrolling when I saw a video about MCP (Model Context Protocol) and it clicked.

The video showed how Claude Desktop can connect directly with external APIs. Several ideas came to mind and I said “I’m going to create a POC about this to see what’s up”.

Why it caught my attention

I think MCP is really useful and can be a revolution in the way LLMs work with backends. Basically it offers a natural query language without being technical.

This tool can help service clients to:

  • Exploit their data naturally
  • Generate automatic summaries
  • Find failures in their systems
  • Map complex information
  • Export specific data

I also think it can work excellently as a super powerful backoffice to facilitate the operation of any company, plus offering fast results and improving user support.

The POC I built

To test the concept, I created a basic test API that handles charges (like a payment system). Nothing complex, just endpoints to create, list and update charges.

The idea was to see if I could really tell Claude “create a charge of $100 for juan@email.com” and have it execute automatically against my test API.

My setup

The flow I built was simple:

  • I tell something to Claude
  • Claude identifies which tool to use
  • A script (MCP Server) translates the command to HTTP requests
  • My test API processes the request

I needed three components:

  1. The test API (basic FastAPI)
  2. An MCP Server that translates Claude commands to HTTP
  3. Configure Claude Desktop to recognize my server

The bumps in the process

First problem: I configured python but on Mac you need python3 or better a bash script.

Second problem: I used relative paths and Claude doesn’t recognize them. You need absolute paths.

Third problem: I installed dependencies globally but Claude was using another Python. Ended up with virtual environment.

Fourth problem: My API expected query parameters but was sending JSON body. Typical documentation error.

How it works in practice

Once working, the flow is pretty natural:

  • “Create a charge of $150 for maria@empresa.com”
  • “Show me all pending charges”
  • “Update Maria’s charge to completed”

Claude maintains context, so it understands references like “Maria’s charge”.

What got me excited about the experiment

For end customers: Imagine giving your users the ability to query their data with natural language. “Show me my October sales” or “Export all my active customers”.

For backoffice: Instead of navigating complex interfaces, the team can say “Cancel all expired subscriptions from the north region” and have it execute automatically.

For support: Agents can resolve complex cases by talking with the system instead of searching through multiple screens.

The limitations I found

Only for simple operations: For complex workflows a dedicated UI is still better.

No visual confirmation: Claude executes actions immediately.

Technical setup: Requires technical knowledge to implement.

Claude Desktop dependent: Only works there, not on web.

My conclusion from the POC

I think this is a revolution in terms of what we know today. The possibility of democratizing access to complex systems through natural conversation completely changes the user experience.

But it’s not all roses. There’s a very important issue to validate: data security and how vulnerable this type of integration can be.

I’m going to explore that in part 2 of this topic.

The complete POC code is at mcp-poc if you want to replicate the experiment.

Comments

You May Also Enjoy