MCP & Agent Learning Journey

Building real AI agent projects from scratch to understand MCP, multi-agent orchestration, and A2A systems.

Week 1 - Orchestrator Loop from Scratch

Goal: Understand how an agent loop works before using any framework.

Built a raw Python orchestrator using only the Groq API. No LangChain. No CrewAI. Every line written and understood manually.

What I learned:

• The orchestrator is just a Python while loop

• LLMs return JSON strings - json.loads converts them to dicts

• Tool descriptions are instructions the LLM reads as text

• Context window fills up - messages list must be managed

Files:

• calculator_agent.py - multi-tool agent with add, multiply, subtract

Related MCP server: ExpenseTracker

Week 2 - Building an MCP Server from Scratch

Goal: Understand MCP as a protocol by building a server, not just using one.

Built a personal finance MCP server with 3 tools. Connected to Claude Desktop. Claude calls my Python functions live.

What I learned:

• MCP server is a separate process, not a library

• list_tools() advertises what the server can do

• call_tool() executes functions and returns structured TextContent

• Claude Desktop spawns the server automatically via config

• Tool descriptions are contracts between the LLM and your code

Tools built:

• compound_interest - future value of an investment

• loan_emi - monthly loan payment calculator

• savings_goal - time to reach a savings target

Files:

• finance_mcp_server.py

Week 3 - Coming Soon

Multi-agent system with CrewAI where each agent uses MCP tools.

Stack

• Python 3.13

• Groq API (Llama 3.3 70B)

• MCP SDK

• Claude Desktop