MCP verification is the new gold in the AI coding era

This code goes with the tutorial MCP Verification Is the New Gold in the AI Coding Era. It shows proper MCP use in Python. The full tutorial text is in Proper_MCP_Use_for_Python_Code.docx.

The official MCP Python SDK is async-only. When you use it in the usual way, async spreads to every function. Also, a stopped VS Code Debug Console cannot send requests to the MCP server.

You cannot fix a bug that you do not know about. To find these bugs, you must stop the program, examine each value and test the result by hand. This is step-by-step verification. This repository shows a different design that makes this possible: threads outside, asyncio in a box. One background thread (the messenger) owns the event loop and the MCP session. All other code is plain Python with threads. You can set a breakpoint on any line, and you can call a tool from the Debug Console while the program is stopped.

Files

File	What it contains
Proper_MCP_Use_for_Python_Code.docx	The tutorial: why async-only MCP code is difficult to debug, the bridge design, measurements and a checklist
mcp_bridge.py	The bridge: the class SyncMCP and a github() helper for GitHub's MCP server
agent_example.py	A plain, threaded agent that uses the bridge
async_style.py	The same agent, written in the usual all-async style, for comparison
tests/run_examples.py	Runs the three examples against a local fake server
tests/debug_console_check.py	The pause-and-ask test: stops each design under debugpy and calls a tool from the Debug Console
tests/fake_github_server.py	A local stand-in for GitHub's MCP server, with fake data

Requirements

• Python 3.10 or later (tested with 3.11 to 3.13)
• The 2.x MCP Python SDK: mcp 2.2 or later. It installs httpx2 automatically. The code does not work with the 1.x SDK.
• debugpy, only for tests/debug_console_check.py

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Run the agent against GitHub

The bridge uses GitHub's remote MCP server with read-only tools. You need a GitHub token.

set GITHUB_TOKEN=<your token>
python agent_example.py
python agent_example.py --serial

With --serial, the jobs run one after the other. Then you can follow one job from start to end in the debugger.

Run the checks (no token necessary)

python tests/run_examples.py
python tests/debug_console_check.py

run_examples.py runs all three examples and examines the bridge: errors, cancellation, threads and clean shutdown. debug_console_check.py does the pause-and-ask test for both designs. With the bridge, the Debug Console gets an answer while the program is stopped. With the all-async design, await never gets an answer.

Try the pause-and-ask test in VS Code

1. Open agent_example.py.
2. Set a breakpoint on the line if result.is_error:.
3. Press F5. VS Code stops in a worker thread.
4. In the Debug Console, type gh.call_tool("get_me", {}). You get an answer.
5. Look at the thread list. It shows MainThread, mcp-loop and the worker threads.

License

Copyright (c) 2026 Sander Stepanov. All rights reserved. No license is granted: you may not use, copy, modify or share this material without written permission. Read LICENSE.