gdb-mcp

# gdb-mcp Model Context Protocol server that wraps `gdb` so an LLM can drive live debugging sessions. The server exposes tools for starting a session on a binary, attaching to an existing PID, running ad-hoc commands, batching commands, checking status, and shutting sessions down. > Requires `gdb` on the host and the `mcp` python package (v1+). ## Setup You can install either inside a virtualenv (recommended) or directly: ```bash # optional but recommended python -m venv .venv . .venv/bin/activate pip install -e . # or: pip install . ``` ## Running the MCP server You can start the server directly: ```bash gdb-mcp # or python -m gdb_mcp.server ``` The server uses stdio for transport (the default for most MCP clients). Point your MCP client configuration at the command above. ### One-shot client configuration Run `gdb-mcp --install` to add the server to any detected MCP-aware clients without editing config files by hand. The installer currently knows how to update: - Codex CLI (`~/.codex/config.toml`) - Claude Desktop configs (common Linux/macOS paths) - Cursor / Windsurf MCP config files (if present) Use `--install-command /custom/path/to/gdb-mcp` or `--install-args ...` if you need to override what gets written. ## Exposed tools - `start_binary(binary_path, args=None, cwd=None, load_init=True, start_timeout=30.0, prompt=None, force_prompt=True)` – launch gdb against a target binary (with optional args) and return a session id and the initial banner. Set `load_init=False` to skip user gdbinit (default is to load it so helpers like GEF/pwndbg remain enabled). Increase `start_timeout` if loading extensions takes longer to reach the first prompt. Use `prompt` to force a specific prompt string (defaults cover `(gdb)`, `(pwndbg)`, and `gef>`). `force_prompt` sets the prompt to the chosen value after startup to avoid timeouts from customized prompts (e.g., pwndbg); set `force_prompt=False` if you need to keep your custom prompt unchanged. -…