DebugMCP: give your AI agent a real debugger

Your AI coding agent can scaffold a function, write tests, and refactor a class — but the moment a bug slips through to runtime, it reverts to reading stack traces and guessing. It cannot set a breakpoint. It cannot step through execution. It cannot inspect local variables while the code is actually running. Every debug cycle becomes a loop of "edit, run, read logs, guess again" — with the AI doing the guessing, not the debugging.

DebugMCP v1.1.4 closes that gap. It is a VS Code extension from Microsoft DevLabs that runs an MCP (Model Context Protocol) server inside the IDE, giving any compatible AI assistant full, programmatic access to the VS Code debugger. 1

What DebugMCP does

DebugMCP exposes 13 MCP tools that map directly to debugger actions your AI can now call autonomously: start_debugging, add_breakpoint, step_over, step_into, step_out, continue_execution, get_variables_values, evaluate_expression, and more. 1

The extension supports 9 languages — Python, JavaScript/TypeScript, Java, C#/.NET, C/C++, Go, Rust, PHP, and Ruby — each backed by the corresponding VS Code language extension and its debug adapter. 1 On the AI side, it auto-detects and registers with 8 assistants: GitHub Copilot, GitHub Copilot CLI, Cline, Cursor, Codex, Windsurf, Roo Code, and Antigravity — and works with any other MCP-compatible assistant. 2

All communication stays on localhost:3001. No code, variable values, or debug data leaves the machine. 2

The architecture is straightforward: the AI sends MCP calls over Streamable HTTP to DebugMCP's local server, which translates them into VS Code's Debug Adapter Protocol (DAP) commands that drive the actual language-specific debugger.

DebugMCP architecture: AI Agent → MCP/Streamable HTTP → DebugMCPServer → DebuggingHandler → VS Code Debug API

Image from: microsoft/DebugMCP — GitHub

Getting started: three steps

Step 1 — Install. Run in your terminal:

code --install-extension ozzafar.debugmcpextension

Or search for "DebugMCP" directly on the VS Code Marketplace. The extension activates automatically on install and starts the MCP server.

Step 2 — Register with your AI assistant. DebugMCP scans for installed AI assistants and offers to register itself automatically. Accept the prompt, or open the Command Palette and run DebugMCP: Show Agent Selection Popup to trigger it manually. 2

For GitHub Copilot, the registration adds the following to your VS Code settings.json:

{
  "mcp": {
    "servers": {
      "debugmcp": {
        "type": "http",
        "url": "http://localhost:3001/mcp",
        "description": "DebugMCP - Multi-language debugging support"
      }
    }
  }
}

Step 3 — Enable auto-approval. In your AI assistant's tool settings, enable auto-approval for all DebugMCP tools. Without it, every step_over and get_variables_values call will pause to ask for confirmation, breaking the flow of a multi-step debugging session. 2

No launch.json is required. DebugMCP auto-generates a debug configuration from the file extension of the file you ask the AI to debug.

Demo: AI debugs a Python bug end-to-end

Here is a minimal Python file with a division bug. The function returns 2 when it should return 2.5:

# calculator.py
def divide(a, b):
    result = 0
    result = a // b   # Bug: integer division instead of float
    return result

print(divide(5, 2))   # Expected: 2.5  |  Actual: 2

Ask your AI assistant (Copilot, Cline, Cursor — any of the supported ones): "Debug calculator.py — set a breakpoint at the division line and inspect why divide(5, 2) returns 2 instead of 2.5."

The AI then executes the following sequence via DebugMCP tools:

start_debugging(fileFullPath: "/path/to/calculator.py") — launches a Python debug session via debugpy
add_breakpoint(lineContent: "result = a // b") — sets the breakpoint by matching the line text 1
continue_execution() — runs the program until the breakpoint is hit
get_variables_values(scope: "local") — reads the local scope: a=5, b=2, result=0
evaluate_expression(expression: "a / b") — evaluates true float division: returns 2.5
Reports: "// performs integer division; change it to / to get the float result."

No log-reading, no guessing.

DebugMCP animated demo — AI assistant setting breakpoints, stepping through code, and inspecting variables

Image from: microsoft/DebugMCP — GitHub

The prerequisite for Python is the Python extension (ms-python.python) — DebugMCP uses VS Code's built-in debugpy adapter, so no separate debugger setup is needed.

Limitations worth knowing before you ship it

DebugMCP is production-ready for most use cases but has a few active bugs as of v1.1.4:

step_into silently degrades to step_over (Issue #71, opened May 10, 2026): after any step-triggered pause, calling step_into on a method call behaves identically to step_over. The failure is silent — no error, no warning. 3 Workaround: call get_stack_trace between every step.
WSL / Linux C/C++ debugging breaks (Issue #70, May 9, 2026): DebugMCP infers a Windows-style .exe target and ignores the program field in an existing launch.json. Manual F5 still works. 4
One debug session at a time: no multi-session support. Running two Claude Code instances (VS Code extension + desktop CLI) against the same DebugMCP server corrupts the server state and requires a window reload. 5
VS Code only: DebugMCP cannot run as a standalone MCP server outside of VS Code. 5

Install and community reception

Marketplace: ozzafar.debugmcpextension 1
CLI: code --install-extension ozzafar.debugmcpextension
License: MIT
Requirements: VS Code 1.104.0+; Windows, macOS, or Linux 1

The GitHub repository has 349 stars and 30 forks as of May 17, 2026 — the highest community signal among the VS Code plugins scanned this week. 2 Gerald Versluis, a Microsoft software engineer working on .NET MAUI and GitHub Copilot, called it out publicly in January 2026:

Gerald Versluis @jfversluis·17w

Controlling your debug sessions through Copilot? Absolutely now you can! Check out this MCP for @code

View on X

正在加载内容卡片…

That tweet drew 245 likes, 46 retweets, and 18,463 views. 6

DebugMCP: Give Your AI Agent a Real Debugger in VS Code