AI coding agents are excellent at writing code, but surprisingly bad at debugging it. In most cases they add print statement, rerun the code and repeat until something works (or doesn’t)..

I built DebugMCP to let AI agents use actual debugging tools inside VS Code - set breakpoints, inspect variables and step through execution.

The idea is simple: give agents the same tools developers rely on.

In my experiments, this changes how bugs get fixed: instead of trial-and-error, the agent can follow the program state.

This is still early, and I’m trying to understand: - does this meaningfully improve real workflows? - what debugging capabilities are actually missing for agents?

Would love feedback, especially from people using Copilot / Claude / Cursor.

Repo: https://github.com/microsoft/DebugMCP