To explore basic building blocks of MCP and intricacies of tool calling within MCP, from scratch. Keeping it as simple as possible without handling many edge cases. Understand how to build both MCP Server and clients.
🚨 Most of the technolgies used in this repo like FastMCP, tool calling in anthropic and MCP itself are very recent. So there are some features in these technologies that are either not available or buggy. But the main goal is a simple exploration without using cursor or Claude Desktop 🚨
Is built using FastMCP, due to its simplicty and automatic error handling. Points to remember:-
Makes a connection to MCP server and should have access to LLM Would higly suggest to explore and inspect the objects returned when the model uses a tool and a genertic reponse Try to stress test giving numbers in different format. Eg:- Add 2 and 3 v/s Add two and three. Eg:- Try to use ambiguous and large numbers:- Add seven hundred thousand trillion with three hundred billion
uv sync
source .venv/bin/activate
Go to fastmcp_calculator_server folder and run
mcp dev server.py

Add your anthropic API key to .env file
Go to mcp_calculator/mcp_compliant_client and run
python client.py <path to server.py>
More about this on medium.
