Simple MCP Server + Client (Python)

A minimal, heavily commented example of the Model Context Protocol (MCP).

How it fits together

client.py                                server.py
---------                                ---------
starts "python server.py"  ───────────▶  process starts, waits on stdin
initialize                 ───stdin──▶   reply with name + capabilities
tools/list                 ───stdin──▶   list of tools + JSON schemas
tools/call add {a:5,b:7}   ───stdin──▶   runs add(5, 7)
                           ◀──stdout───  result: 12
close connection           ───────────▶  process exits

All messages are JSON-RPC 2.0 sent over the server’s stdin/stdout (“stdio transport”).

Setup & run

python -m venv .venv

.venv\Scripts\pip install -r requirements.txt

.venv\Scripts\python client.py

Use the server from Claude Code

claude mcp add demo-server -- C:\Users\ab845312\Downloads\github\mcp\.venv\Scripts\python.exe C:\Users\ab845312\Downloads\github\mcp\server.py

Chat with a local LLM (chat_client.py)

client.py uses hard-coded inputs because there’s no AI in the loop. chat_client.py adds one: you type a question, a local open-source model (Qwen3 4B, run via Ollama) decides which tool to call, the client runs it on the MCP server, and the model answers using the result.

Prerequisites: install Ollama, then pull the Qwen3 model (it supports tool calling):

ollama pull qwen3:4b

Run it:

.venv\Scripts\python chat_client.py

Use a different model:

.venv\Scripts\python chat_client.py --model mistral

Example session:

You: What is 12.3 times 4?
  [tool] multiply({'a': 12.3, 'b': 4})
  [tool] -> 49.2
AI: 12.3 times 4 equals 49.2.

Notes:

• Only the server’s tools are given to the model (not resources or prompts).
• Small models sometimes skip a tool and guess, or write the tool call as plain text. chat_client.py includes a small workaround for the latter (recover_text_tool_calls).