WinDbg Agent

WinDbg/CDB extension that exposes the debugger via HTTP and MCP servers, letting external AI agents (Claude Code, Copilot, etc.) control debug sessions programmatically.

Building

Prerequisites

• Windows 10/11
• Visual Studio 2022 with C++ workload (2019 also works — see below)
• CMake 3.20+
• Windows SDK (for dbgeng.h)

Build Steps

# Clone with submodules
git clone --recursive https://github.com/0xeb/windbg-agent.git
cd windbg-agent

# Build for x64 (most common)
cmake --preset x64
cmake --build --preset x64

# Build for x86 (32-bit targets)
cmake --preset x86
cmake --build --preset x86

Output:

• x64: build-x64/Release/windbg_agent.dll and windbg_agent.exe
• x86: build-x86/Release/windbg_agent.dll and windbg_agent.exe

Alternative: No Visual Studio 2022

If you have a different Visual Studio version or just the Build Tools, skip the presets and specify the generator manually:

# Visual Studio 2019 — x64
cmake -B build-x64 -G "Visual Studio 16 2019" -A x64
cmake --build build-x64 --config Release

# Visual Studio 2019 — x86 (32-bit targets)
cmake -B build-x86 -G "Visual Studio 16 2019" -A Win32
cmake --build build-x86 --config Release

# Ninja (works with any MSVC version — run from Developer Command Prompt)
cmake -B build-x64 -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build-x64

Ninja x86 note: For 32-bit Ninja builds, open the x86 Native Tools Command Prompt (instead of x64) so cl.exe targets Win32.

Usage

Loading the Extension

1. Open WinDbg and load a dump file or attach to a process
2. Load the extension:

.load C:\path\to\windbg_agent.dll

Note: The extension DLL architecture must match the target. Use the x86 build when debugging 32-bit processes, and the x64 build for 64-bit processes.

Commands

Command	Description
!agent help	Show help
!agent version	Show version
!agent http [bind_addr]	Start HTTP server (port auto-assigned)
!agent mcp [bind_addr]	Start MCP server for MCP-compatible clients

HTTP Server

Start an HTTP server to let external tools execute debugger commands:

!agent http                  # localhost only (default)
!agent http 0.0.0.0          # all interfaces (no auth — use with care)

The server prints the auto-assigned URL, endpoints, and usage examples. The output is also copied to the clipboard.

Endpoints:

Method	Path	Description
POST	/exec	Execute a debugger command
GET	/status	Server status
POST	/shutdown	Stop the server

Examples:

# Execute a debugger command
curl -X POST http://127.0.0.1:<port>/exec \
  -H "Content-Type: application/json" \
  -d '{"command": "kb"}'

# Python
import requests
r = requests.post('http://127.0.0.1:<port>/exec', json={'command': 'kb'})
print(r.json()['output'])

Response format:

{"output": "...", "success": true}

CLI Tool

A standalone HTTP client (windbg_agent.exe) for scripting against a running HTTP server:

windbg_agent.exe --url=http://127.0.0.1:<port> exec "kb"
windbg_agent.exe --url=http://127.0.0.1:<port> status
windbg_agent.exe --url=http://127.0.0.1:<port> shutdown

The URL defaults to http://127.0.0.1:9999 or the WINDBG_AGENT_URL environment variable.

MCP Server

Start an MCP (Model Context Protocol) server for MCP-compatible AI clients:

!agent mcp                   # localhost only (default)
!agent mcp 0.0.0.0           # all interfaces

Exposes a single dbg_exec tool. Add to your MCP client config:

{
  "mcpServers": {
    "windbg-agent": {
      "url": "http://127.0.0.1:<port>/sse"
    }
  }
}

Screenshots

Start !agent http to let external tools control WinDbg:

The CLI executes commands remotely:

Claude Code controlling WinDbg via HTTP server — analyzing a crash:

Author

Elias Bachaalany (@0xeb)

Pair-programmed with Claude Code and Codex.

License

MIT