An MCP (Model Context Protocol) server that exposes every pybag Windows debugger function as a native MCP tool. It gives any MCP-compatible client (Claude Desktop, Claude Code, Cowork, OpenAI Codex CLI, Cursor, and custom agents) full control over user-mode processes, kernel sessions, and crash dump analysis — all through typed tool calls with structured JSON responses.

• Windows only — pybag requires Microsoft Debugging Tools for Windows
• Python 3.10+
• Microsoft Debugging Tools for Windows (part of the Windows SDK)

git clone https://github.com/your-username/windbg-mcp.git
cd windbg-mcp

pip install pybag mcp

Download the Windows SDK and select Debugging Tools for Windows during setup: https://developer.microsoft.com/en-us/windows/downloads/windows-sdk/

The server runs as a local stdio process. All clients below launch it the same way — python <path-to>/windbg_mcp.py — but each has its own config format.

Edit the Claude Desktop configuration file and add the windbg-mcp entry:

Config file location:

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "windbg-mcp": {
      "command": "python",
      "args": ["C:\\path\\to\\windbg-mcp\\windbg_mcp.py"]
    }
  }
}

Restart Claude Desktop. All 55 debugger tools will appear automatically.

Run the following command once to register the server. Claude Code stores the entry in its own MCP config and makes the tools available in every subsequent session.

claude mcp add windbg-mcp python C:\path\to\windbg-mcp\windbg_mcp.py

To verify the server was registered:

claude mcp list

To remove it later:

claude mcp remove windbg-mcp

There are two ways to add WinDbg MCP to Cowork: via JSON configuration (quick) or by installing it as a .mcpb plugin bundle (portable, shareable).

1. Open the Claude desktop app and go to Settings → MCP Servers.
2. Click Add Server and paste the following:

{
  "windbg-mcp": {
    "command": "python",
    "args": ["C:\\path\\to\\windbg-mcp\\windbg_mcp.py"]
  }
}

3. Save and restart Cowork. The tools will be available in your next session.

A .mcpb file is a zip archive of the plugin directory that Cowork can install directly. This is the recommended approach when sharing the server with a team or across machines.

Step 1 — Build the .mcpb file

From the root of the cloned repository, run:

powershell -Command "Compress-Archive -Path '.\*' -DestinationPath 'windbg-mcp.zip'; Rename-Item 'windbg-mcp.zip' 'windbg-mcp.mcpb'"

This creates windbg-mcp.mcpb in the current directory, bundling windbg_mcp.py, manifest.json, and any other project files.

Step 2 — Install in Cowork

1. Open the Claude desktop app.
2. Go to Settings → Plugins (or Extensions).
3. Click Install Plugin and select windbg-mcp.mcpb.
4. Cowork reads manifest.json from the bundle, registers the MCP server, and makes all tools available immediately — no manual path configuration required.

The manifest.json bundled in this repo is already configured correctly:

{
  "manifest_version": "0.2",
  "name": "windbg-mcp",
  "version": "1.0.0",
  "description": "WinDbg MCP — full Windows debugger control via MCP tools",
  "server": {
    "type": "python",
    "entry_point": "windbg_mcp.py",
    "mcp_config": {
      "command": "python",
      "args": ["${__dirname}/windbg_mcp.py"]
    }
  }
}

${__dirname} is resolved at install time to the directory where Cowork unpacked the bundle, so you do not need to hard-code any paths.

Add the server to your Codex CLI configuration file. The file is typically located at ~/.codex/config.json (Linux/macOS) or %USERPROFILE%\.codex\config.json (Windows).

{
  "mcpServers": {
    "windbg-mcp": {
      "command": "python",
      "args": ["C:\\path\\to\\windbg-mcp\\windbg_mcp.py"]
    }
  }
}

Once saved, start a new Codex session. The WinDbg tools will be available for the model to call.

1. Open Cursor → Preferences → Cursor Settings.
2. Navigate to the MCP tab.
3. Click Add new global MCP server and use this configuration:

{
  "windbg-mcp": {
    "command": "python",
    "args": ["C:\\path\\to\\windbg-mcp\\windbg_mcp.py"]
  }
}

4. Save. Cursor will connect to the server on its next Composer session.

Add the following to your ~/.continue/config.json (or the workspace-level .continue/config.json):

{
  "experimental": {
    "modelContextProtocolServers": [
      {
        "transport": {
          "type": "stdio",
          "command": "python",
          "args": ["C:\\path\\to\\windbg-mcp\\windbg_mcp.py"]
        }
      }
    ]
  }
}

Reload the Continue extension. The 55 debugger tools will appear in the tool list.

If you are building your own agent or automation pipeline, connect to WinDbg MCP over the standard MCP stdio transport. The server speaks JSON-RPC 2.0 over stdin/stdout.

import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

server_params = StdioServerParameters(
    command="python",
    args=[r"C:\path\to\windbg-mcp\windbg_mcp.py"],
)

async def main():
    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()

            # List all available tools
            tools = await session.list_tools()
            print([t.name for t in tools.tools])

            # Load a crash dump
            result = await session.call_tool(
                "load_dump",
                arguments={"path": r"C:\crashes\crash.dmp"},
            )
            print(result.content)

            # Read 64 bytes at RSP
            result = await session.call_tool(
                "read_mem",
                arguments={"addr": "0x00000000001FF000", "size": 64},
            )
            print(result.content)

asyncio.run(main())

import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";

const transport = new StdioClientTransport({
  command: "python",
  args: ["C:\\path\\to\\windbg-mcp\\windbg_mcp.py"],
});

const client = new Client({ name: "my-agent", version: "1.0.0" }, {});
await client.connect(transport);

// Call a tool
const result = await client.callTool({
  name: "load_dump",
  arguments: { path: "C:\\crashes\\crash.dmp" },
});
console.log(result.content);

await client.close();

from langchain_mcp_adapters.tools import load_mcp_tools
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

server_params = StdioServerParameters(
    command="python",
    args=[r"C:\path\to\windbg-mcp\windbg_mcp.py"],
)

async def get_tools():
    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()
            return await load_mcp_tools(session)

The server communicates via newline-delimited JSON-RPC 2.0 messages. You can drive it from any language by writing to the process's stdin and reading from stdout:

→ {"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"my-client","version":"1.0"}}}
← {"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2024-11-05","capabilities":{...},"serverInfo":{"name":"WinDbg MCP","version":"1.0.0"}}}

→ {"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"load_dump","arguments":{"path":"C:\\crashes\\crash.dmp"}}}
← {"jsonrpc":"2.0","id":2,"result":{"content":[{"type":"text","text":"{\"status\": \"ok\", ...}"}]}}

create — Launches a new process under the debugger. Set initial_break=True (default) to break at the process entry point.

attach — Attaches to a running process. Provide either pid (integer) or name (process filename). Do not provide both.

kernel_attach — Connects to a remote kernel debugger. connect_string uses KD syntax, e.g. "net:port=55000,key=1.2.3.4".

load_dump — Opens a .dmp file for post-mortem analysis. Returns the crash address and nearest symbol immediately.

connect — Connects to a process server for remote user-mode debugging. options uses DbgEng connection syntax, e.g. "tcp:server=192.168.1.10,port=5555".

go — Resumes execution and blocks until the next debug event (breakpoint, exception, or timeout). Returns the new RIP and any captures collected while running.

step_into — Steps into the next instruction, following calls into called functions.

step_over — Steps over the next instruction, treating calls as a single step.

step_out — Runs until the current function returns.

goto — Runs until a specific symbol or hex address is reached, e.g. "Kernel32!ExitProcess" or "0x7fff12340000".

trace — Performs N single-step iterations and records each instruction visited.

bp — Sets a software (code) breakpoint at a symbol or address.

• expr: symbol ("ntdll!NtCreateFile") or hex address ("0x7ff800001234")
• capture: when true (default), automatically saves full state — registers, stack, memory — to the capture buffer each time this breakpoint fires
• action: "go" (default) continues execution after capture; "break" halts
• oneshot: removes the breakpoint after it fires once
• passcount: fires only after N passes through the location

hw_bp — Sets a hardware / data breakpoint (watchpoint).

• addr: hex address to watch
• size: watch width in bytes — 1, 2, 4, or 8 (default 4)
• access: "e" execute, "w" write (default), "r" read/write
• capture, action, oneshot: same semantics as bp

list_bps — Returns all currently active breakpoints with their IDs, expressions, types, and settings.

remove_bp / enable_bp / disable_bp — Manage breakpoints by the id returned from bp or hw_bp.

Breakpoints with capture: true (the default) automatically save a full debugger snapshot every time they fire. The snapshot includes all registers, the call stack, 64 bytes of stack memory at RSP, and 32 bytes of code at RIP. Snapshots accumulate in a buffer and can be retrieved at any time with get_captures.

get_captures — Returns all captures collected since the last clear_captures. Each capture contains:

• registers — all register values as {name: "0x..."} hex strings
• rip — instruction pointer at the moment of capture
• symbol_at_rip — nearest symbol to RIP
• instruction — disassembly of the instruction at RIP
• stack — top 10 call stack frames with addresses and return addresses
• context_memory.stack_at_rsp — 64 bytes at RSP as hex, formatted, and ASCII
• context_memory.code_at_rip — 32 bytes at RIP as hex and formatted

clear_captures — Clears the capture buffer. Useful before starting a new run.

capture_state — Takes an immediate on-demand snapshot of the current state. Use this when already broken in, rather than waiting for a breakpoint to fire.

read_mem — Reads size raw bytes from addr. Returns the data as hex (compact), formatted (space-separated bytes), and ascii (printable characters, . for non-printable).

write_mem — Writes bytes to memory. data is a hex string — spaces and \x prefixes are stripped automatically, e.g. "90909090", "\\x90\\x90\\x90\\x90", or "90 90 90 90".

read_ptr — Reads count consecutive pointer-sized values (4 bytes on 32-bit, 8 bytes on 64-bit) starting at addr.

poi — Dereferences a single pointer at addr (pointer-of-interest).

read_str — Reads a null-terminated string. Set wide=true for UTF-16LE (Windows WCHAR).

dump_mem — Formatted dword/pointer dump, equivalent to dd/dp in WinDbg.

mem_info — Returns the memory region properties for the page containing addr: base address, size, type, state, and protection flags.

mem_list — Lists all virtual memory regions in the target process address space.

get_regs — Returns every available register as {name: "0x..."}. The exact set depends on the target architecture (x86 vs x64).

get_reg — Returns a single register, e.g. name="rax", name="eflags".

set_reg — Overwrites a register. value accepts hex strings ("0x1234") or decimal integer strings.

get_pc — Returns the instruction pointer with symbol resolution and the decoded instruction text at that address.

get_sp — Returns the current stack pointer value.

resolve — Resolves a symbol name to its virtual address. Use Module!Function format, e.g. "Kernel32!WriteFile", "ntdll!NtCreateFile".

find_symbols — Wildcard symbol search, e.g. "ntdll!*Alloc*", "kernel32!*File*". Returns all matching symbol strings.

addr_to_symbol — Reverse-resolves a virtual address to the nearest symbol name.

disasm — Disassembles count instructions starting at addr. Defaults to the current RIP if no address is given.

whereami — Returns a human-readable description of the module, function, and offset at the given address.

list_modules — Lists all modules loaded in the target, with their base address and size.

module_info — Returns the entry point and section list (name, virtual address, size) for a specific module, e.g. "kernel32.dll", "ntdll.dll".

get_exports — Returns the full export table of a module as a list of strings.

get_imports — Returns the full import table of a module as a list of strings.

list_threads — Lists all threads in the target process.

get_thread — Returns the currently active thread context.

set_thread — Switches the active thread context by thread ID (from list_threads).

get_stack — Returns the call stack as structured data. Each frame includes the instruction address, return address, and frame pointer.

get_teb — Returns the address of the Thread Environment Block for the current thread.

get_peb — Returns the address of the Process Environment Block.

get_handles — Lists all open handles in the target process.

get_bitness — Returns 32 or 64 depending on the target architecture.

raw — Executes any WinDbg command string and returns the output as text. Use this as an escape hatch for anything not covered by the other tools:

raw(cmd="!heap -stat")
raw(cmd="dt _PEB @$peb")
raw(cmd="!locks")
raw(cmd="lm")
raw(cmd="!address @rsp")

1. create(path="C:/target/vuln.exe", args="exploit_input.bin")
2. bp(expr="vuln!processInput+0x2A", action="https://github.com/break")
3. go(timeout=15000)
4. get_captures()

In get_captures, inspect captures[0].registers.rip:

• "0x4141414141414141" — you control RIP with 'A' bytes
• Any value matching your pattern — controlled
• A valid-looking address — crash but not yet controlled

Check captures[0].context_memory.stack_at_rsp.formatted to see padding, return addresses, or shellcode bytes on the stack.

1. load_dump(path="C:/crashes/crash.dmp")
2. get_regs()             → full register state at crash time
3. get_stack(frames=30)   → call stack at crash
4. get_sp()               → read RSP value
5. read_mem(addr=<rsp>, size=64) → stack contents
6. disasm()               → instructions at the crash address

1. attach(name="target.exe")
2. hw_bp(addr="0x1001F000", size=8, access="w", action="https://github.com/break")
3. go()
4. get_captures()                            → see what wrote to the spray address
5. read_mem(addr="0x1001EFC0", size=128)     → surrounding memory context

1. create(path="C:/target/target.exe")
2. resolve(name="kernel32!WriteFile")   → record base address
3. terminate()
4. create(path="C:/target/target.exe")
5. resolve(name="kernel32!WriteFile")   → compare: changed = ASLR on, same = ASLR off

1. kernel_attach(connect_string="net:port=55000,key=1.2.3.4")
2. list_modules()                        → all loaded kernel modules
3. module_info(name="ntoskrnl.exe")      → entry point and sections
4. raw(cmd="!process 0 0")              → list all processes from kernel context
5. raw(cmd="!pcr")                       → processor control region

1. attach(pid=1234)
2. list_threads()          → all thread IDs
3. set_thread(id=2)        → switch context
4. get_stack(frames=20)    → call stack for that thread
5. get_regs()              → registers for that thread
6. get_teb()               → TEB address

Symbol path — If symbol resolution returns no results, configure the Microsoft symbol server:

raw(cmd=".sympath srv*C:\\symbols*https://msdl.microsoft.com/download/symbols")
raw(cmd=".reload")

Timeout tuning — go() defaults to 30 seconds. For targets that run longer before hitting a breakpoint:

go(timeout=120000)   # 2 minutes
go(timeout=300000)   # 5 minutes

Address format — All addr parameters accept hex strings ("0x1234abcd", "7fff12340000") or plain integers. The 0x prefix is optional for hex values.

Shellcode verification — After a capture, use read_mem and disasm on the address where your shellcode should land. If disasm shows your intended instructions, the payload arrived intact.

After terminate or detach — All captures and breakpoints are cleared automatically. Call create or attach to begin a new session.

capture_state vs get_captures — Use capture_state for an on-demand snapshot when already stopped at a breakpoint. Use get_captures to retrieve state that was automatically saved each time a breakpoint fired during a go call.

Kernel raw commands — Common kernel debugging extensions that work well through raw:

raw(cmd="!process 0 0")       → list all processes
raw(cmd="!thread")            → current thread details
raw(cmd="!irql")              → current IRQL
raw(cmd="!pcr")               → processor control region
raw(cmd="!pte <addr>")        → page table entry for an address
raw(cmd="dt nt!_EPROCESS @$proc")  → dump EPROCESS structure

MIT