GDB Lite MCP

A small TypeScript MCP server for driving GDB sessions from LLM agents.

GDB Lite MCP intentionally exposes only a thin set of primitives around native GDB. Agents can spawn sessions, execute GDB commands, interrupt hung programs, and close sessions while still using normal GDB features such as breakpoints, watchpoints, command lists, Python snippets, core files, attach, and remote targets.

Requirements

• Node.js 20 or newer

• GDB available on PATH

• A native compiler such as gcc if you want to build the repository scenarios

Related MCP server: gdb and rr Debugging

Install

npm install
npm run build

Run the server locally:

npm start

The package also exposes a gdb-lite-mcp binary after it has been built.

The npm package is intentionally limited to the runtime server, debug guide, and debugging Skill. The eval/scenarios/ directory contains repository development assets; clone this repository if you want to run them locally.

MCP Configuration

Point your MCP client at the published package via npx:

{
  "mcpServers": {
    "gdb-lite": {
      "command": "npx",
      "args": ["-y", "gdb-lite-mcp"]
    }
  }
}

For repository-local evaluation, eval/run_eval.py writes a temporary opencode.json that starts the same server from the built dist/index.js.

Runtime environment variables:

• GDB_LITE_GDB_PATH: GDB executable path. Defaults to gdb.

• GDB_LITE_MAX_SESSIONS: maximum live sessions. Defaults to 8; must be a positive integer.

• GDB_LITE_MAX_INTERNAL_BUFFER_CHARS: per-session retained output buffer. Defaults to 4194304; must be a positive integer.

• GDB_LITE_AUTO_INIT: set to 0, false, no, or off to disable automatic startup defaults.

Invalid runtime configuration values fail fast when the server starts.

Tools

The server registers these MCP tools:

gdb_exec and gdb_interrupt return structured state such as completion_reason (completed, timeout, or exited), at_prompt, command_pending, needs_interrupt, timed_out, truncated, and byte counts. Use this metadata to avoid stacking commands behind a still-running inferior. Calls on the same session are not queued; concurrent gdb_exec or gdb_interrupt requests are rejected.

The server also exposes a gdb-lite://debug-guide resource backed by GUIDE.md.

Example Workflow

gdb_spawn({
  "prog_path": "bin/program",
  "work_dir": "/absolute/path/to/debug-workspace"
})

gdb_exec({
  "session_id": "...",
  "command": "break main\nrun\nbt\ninfo locals",
  "timeout": 5
})

gdb_close({
  "session_id": "..."
})

Prefer batching related GDB commands in one tool call. For hangs, let the run time out, call gdb_interrupt, then inspect bt, info threads, and relevant locals before continuing.

Debugging Skill

The skills/gdb-debugging directory contains an agent Skill with focused debugging workflows for:

• crashes

• hangs

• memory corruption

• recursion issues

• wrong results

• GDB Python probes

Agents that support repository-local Skills should read skills/gdb-debugging/SKILL.md before using the MCP tools.

Scenarios

The repository eval/scenarios directory contains small native debugging tasks used to evaluate the MCP server and Skill. It is not included in the npm package.

Build all scenario binaries:

python3 eval/scenarios/build_scenarios.py

Each scenario writes local build artifacts under eval/scenarios/<name>/build/, which is ignored by Git.

Evaluation

Repository-local evaluation prompts and config live under eval/. They are not included in the npm package.

npm run build
python3 eval/run_eval.py --scenario hang-tokenizer

Run outputs are written under eval/runs/, which is ignored because it is a local evaluation result artifact.

Development

npm run build
npm test

Generated artifacts such as dist/, node_modules/, scenario build directories, logs, and local evaluation results are ignored by Git.

License

MIT. See LICENSE.