PwndbgBLOCK
An MCP tool endows AI agent with the capability to debug ELF
Overview
From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.
中文说明
An MCP tool endows AI agent with the capability to debug ELF. This tool is designed for regular CTF challenges, especially pwn ones.
[!IMPORTANT] It's recommended to use a tmux skill and pwndbg skill, and provide AI agent with a tmux session instead. Controlling pwndbg with MCP requires lots of tokens, and it's hard to support debugging with python script. Reference our [0RAYS/codex-pwner]. (Though AI agent tends to use raw tools like reading /proc/mem to get information...) [0RAYS/codex-pwner]: https://github.com/0RAYS/codex-pwner
Quickstart
As currently this repo has not been published on PyPI, clone this repo with
$ git clone https://github.com/RocketMaDev/pwndbg-mcp.git
Then use uv (install it if you don't have it) to pull dependencies and build venv, or install it as a tool.
$ uv tool install pwndbg-mcp $ pwndbg-mcp # or clone pwndbg-mcp and run it with uv manually
[!CAUTION] DON'T EXPOSE YOUR MCP SERVICE TO PUBLIC IF YOU DON'T DO ISOLATION! The toolload_executable,execute_command(GDB command) andeval_to_send_to_processmay lead to remote code execution.
By default starting pwndbg-mcp without args launch a mcp server at localhost:8780 with gdb as main pwndbg binary, /mcp as mcp connection end point, and HTTP streamable as transport. Here is some help information:
usage: main.py [-h] [--transport {stdio,http,sse}] [--host HOST] [--port PORT] [--pwndbg BIN] [--d2dname NAME] [--d2dhost HOST] [--d2dport PORT]
pwndbg-mcp: An MCP tool endows AI agent with the capability to debug ELF
options:
-h, --help show this help message and exit
--transport {stdio,http,sse}, -t {stdio,http,sse}
Transport mode: stdio, http (streamable HTTP, default), or sse
--host HOST,48d967daff42OBSERVED · 2026-10-08Connect
Built from this server's own package name, version and transport as found in its source — not copied from anyone's documentation, so it cannot drift against a page we do not control.
claude mcp add pwndbg-mcp -- uvx pwndbg-mcp
{
"mcpServers": {
"pwndbg-mcp": {
"command": "uvx",
"args": [
"pwndbg-mcp"
]
}
}
}Exposed tools (17)
13 read · 3 write · 1 destructive. Blast radius: 1 tool can delete or overwrite — an agent that can be talked into calling a tool can be talked into calling this one.
| Tool | Risk | Description |
|---|---|---|
backtrace | read | Display program function backtrace |
bins | read | Examine available chunks with pwndbg command bins |
context | read | Display pwndbg section with `subsection` or leave it to null |
debug_control | read | Control tracee running state by step, next or finish, etc. Use this prior |
eval_to_send_to_process | write | Given `statement`, evaluate it in Python and `bytes()` it, |
heap | read | Examine heap with pwndbg |
interrupt_process | read | Interrupt target process through PTY. Equivalent to press Ctrl-C, Ctrl-Z or Ctrl-D |
list_pwndbg_commands | read | If you don |
load_executable | write | Load an executable file into GDB and set up PTY for process I/O. |
procinfo | read | Display current process infomation |
pwndbg_hard_reset | destructive | Hard reset pwndbg status by reset gdb controller. If interrupt can not stop process, |
pwndbg_status | read | Return GDB status and additional GDB messages if some new messages |
read_from_process | read | Read data from the target process through PTY. |
send_to_process | write | Send data to the target process through PTY in raw mode. |
telescope | read | Examine given address `statement` with `count` words, |
vmmap | read | Display current program memory layout or pages match pattern |
xinfo | read | Shows offsets of the specified address from various useful locations |
Trust audit
BLOCKgrade D · trust 69/100 Do not install this without reading the findings. The audit found something that could harm you or your machine.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | FAIL |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | WARN |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- declared (1 observation(s))
- Network
- none-observed
- Shell
- declared (1 observation(s))
- Dependencies
- pinned
- Secrets in source
- none-found
Findings (2)
result = eval(statement)
pwndbg_hard_reset
Gates applied: no_behavioural_pass.
48d967daff42full audit observations/trust-audit/mcp-server/rocketmadev__pwndbg.json · Report an issue / request a re-scanAudit history
Every audit this server has had. A grade with a past is a grade somebody is still checking.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-08 | 48d967daff42 | BLOCK | D | 69 | first audit |
Questions
What is the Pwndbg MCP server?
An MCP tool endows AI agent with the capability to debug ELF
What tools does Pwndbg expose?
17 in total: 13 read-only, 3 that write, and 1 that can delete or overwrite (pwndbg_hard_reset). Every one is listed on this page with its risk.
Is Pwndbg safe to connect to an agent?
No — not without reading the findings first. The audit graded it D (69/100) and found 1 critical or high issue in the source. Each one is listed on this page with the file and line it is on. Separately from the audit: 1 of its tools can destroy data, so scope the token you give it to what you actually need.
What credentials does Pwndbg need?
No credential environment variables were found in its source, so it appears to need none.
How current is this page?
The grade is for one exact copy of the source (48d967daff42), read on 2026-10-08. The repository is watched and re-audited when it changes.