Debug native crashes by handing them to your coding agent over ACP (Agent Client Protocol). Exposes a live LLDB debugger as MCP tools.
acpdbg-debugger defines 9 tools with clear, action-verb names (debugger_command, get_backtrace, step_over, etc.) and reasonable descriptions (avg ~120 chars). All tools have input schemas with type definitions. However, several critical gaps reduce quality: (1) Most parameter descriptions are minimal or absent, e.g., set_breakpoint has 'function', 'file', 'line' params with descriptions but no guidance on mutual exclusivity or required combinations; (2) No output schemas documented, LLMs cannot predict what fields to expect from responses; (3) Error handling is generic ('debugger error: {exc}') with no recovery guidance; (4) No tool annotations (readOnlyHint/destructiveHint) despite clear read vs. write distinction; (5) Control tools (step_over, continue_execution, etc.) lack warnings about irreversible side effects. Naming is strong (verb_noun pattern), but parameter and output documentation lag production baselines.
Resume the program until it next stops (breakpoint, signal, or exit) (LLDB 'continue').
Run a single LLDB command against the live stopped process and return its output. Examples: 'bt', 'frame variable', 'p argc', 'x/16xb ptr', 'image lookup -a 0x...'. Read-only inspection is allowed; commands that resume or mutate the process are rejected.
Return the backtrace of the stopped thread (LLDB 'bt').
Return local variables of the selected frame (LLDB 'frame variable').
Set a one-shot breakpoint at file:line and continue to it.
Set a breakpoint by function name, or by file and line.
No output schemas documented. LLMs cannot predict response structure (fields, types, pagination). Baseline: 100% of A+ tools document return types.
set_breakpoint lacks parameter relationship documentation. 'function' vs 'file+line' are mutually exclusive, but this is not stated. LLMs will pass both, causing ambiguous calls.
Control tools (step_over, step_into, step_out, continue_execution, set_breakpoint, run_to_line) lack destructiveHint annotations and warnings about irreversible side effects. Agents need explicit guidance that these resume/mutate the process.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
Step to the next line, descending into any function call (LLDB 'thread step-in' / 'step').
Run until the current function returns to its caller (LLDB 'thread step-out' / 'finish').
Execute the current source line and stop on the next one, without descending into calls (LLDB 'thread step-over' / 'next').
Error responses are generic ('debugger error: {exc}') with no recovery guidance. Baseline: error responses must tell LLM what to do next. E.g., 'Bridge unavailable. Ensure ACPDBG_BRIDGE_FILE is set or session is active.'
Parameter descriptions for control tools are minimal. E.g., set_breakpoint 'function' param lacks guidance on format (bare name vs qualified path). Baseline: 100% of A+ tool params have descriptions >20 chars.