MCP server that provides debugging capabilities via the Debug Adapter Protocol (DAP), supporting multiple debuggers (Delve for Go, GDB for C/C++/Rust)
The server defines 13 debugging tools with explicit schemas and descriptions in tools.go. Tool naming is mostly clear (debug, breakpoint, continue, step, pause, context, evaluate, info, restart, set-variable, disassemble), though some are unconventional for MCP (e.g., 'clear-breakpoints' uses a hyphen; 'stop' lacks a verb prefix). Descriptions are substantial and explain modes/parameters well for complex tools like 'debug', but several tools lack parameter descriptions in the schema definition visible in the code. The debug tool description is comprehensive (~400 chars), but parameter-level descriptions appear to be incomplete, the input schema shows parameter names and high-level descriptions (e.g., 'Mode: source, binary, core, or attach'), but LLMs need per-parameter guidance in the schema object itself. Error handling is minimal, no recovery guidance, no actionable error messages shown. Responses are not documented; output schemas are inferred from DAP protocol but not explicitly declared for the LLM. The breakpoint and clear-breakpoints tools handle multiple exclusive input patterns (file+line vs function vs all), but this is not clearly documented in the schema structure. No security considerations are visible (e.g., injection via expression evaluation in the 'evaluate' tool).
Set a breakpoint. Provide EITHER file+line OR function name (not both). Examples: {"file": "/path/to/main.go", "line": 42} or {"function": "main.processData"}
Remove breakpoints. Provide 'file' to clear all breakpoints in a file, 'file'+'line' to clear a specific line breakpoint, 'function' to clear a function breakpoint, or 'all': true to clear all breakpoints. Examples: {"file": "/path/to/main.go"} or {"file": "/path/to/main.go", "line": 42} or {"function": "main.processData"} or {"all": true}
Get the current debugging context (stack trace, variables, location).
Continue program execution until the next breakpoint or termination. By default returns a compact stop summary (location only). Set fullContext: true only if you need variables immediately — it saves a separate 'context' call but returns much more data. Leave fullContext false (the default) unless you know you need variables right away. Optionally specify 'to' for run-to-cursor: {"to": {"file": "/path/main.go", "line": 50}} or {"to": {"function": "main.Run"}}
Start a complete debugging session. Modes: 'source' (compile & debug), 'binary' (debug executable), 'core' (debug core dump), 'attach' (connect to process). Debugger selection (via 'debugger' parameter): - 'delve' (default): For Go programs only. Requires dlv to be installed. - 'gdb': For C/C++/Rust and other compiled languages. Requires GDB 14+ with native DAP support (gdb -i dap). GDB does not support 'source' mode; compile your program with debug symbols (gcc -g -O0) and use 'binary' mode. Choose the debugger based on the language of the program being debugged: use 'delve' for Go, use 'gdb' for C/C++/Rust. By default, when stopped at a breakpoint returns a compact stop summary (location only). Set fullContext: true only if you need variables immediately — leave it false unless you plan to call 'context' right after anyway.
Output schemas are not documented for any tool. LLMs cannot know what fields 'context' returns, whether 'evaluate' returns a value or an error, or what structure 'info' provides. This forces LLMs to guess or hallucinate field names for downstream processing.
Mutually exclusive input patterns are not clearly documented. 'breakpoint' accepts (file+line) OR (function), and 'clear-breakpoints' accepts (file), (file+line), (function), or (all), but the schema doesn't enforce or explain these constraints, causing LLMs to pass ambiguous or conflicting parameters.
Parameter 'mode' in 'debug' and 'step' tools should be enums, not free-form strings. The description lists valid values ('source', 'binary', 'core', 'attach' for debug; 'over', 'in', 'out' for step), but JSON Schema lacks enum constraints, allowing LLMs to pass invalid values like 'start' or 'skip'.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 56 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 36 | - | v1 |
Show disassembled code at the current location (if supported by the debugger).
Evaluate an expression in the current debugging context.
Get information about the debugging session (current state, capabilities).
Pause a running program at the current execution point.
Restart the debugging session (if supported by the debugger).
Set a variable value in the current debugging context (if supported by the debugger).
Step through code one line at a time. By default returns a compact stop summary (location only). Set fullContext: true only if you need variables immediately — it saves a separate 'context' call but returns much more data. Leave fullContext false (the default) unless you know you need variables right away. Modes: 'over' (execute current line, step over function calls), 'in' (step into function calls), 'out' (run until current function returns).
End the debugging session. By default terminates the debuggee. Pass detach=true to detach without killing the process (leaves it running); detach requires adapter support.
No error handling guidance. Tools provide no recovery hints if the debugger is not installed, a breakpoint fails to set, or an expression evaluation is invalid. LLMs have no actionable next steps on failure.
Security: The 'evaluate' tool accepts arbitrary expressions in the current debugging context. No validation, sandboxing, or permission checks are documented. An adversary could inject code via prompt injection. The tool description lacks any warning about expression safety.
Tool naming inconsistency: 'clear-breakpoints' uses kebab-case; all others use lowercase with no separator or hyphenation. Inconsistent naming reduces discoverability and clarity for LLMs.
Parameter descriptions in the 'debug' tool are inline text rather than per-parameter schema annotations. LLMs parse schema objects more reliably than prose descriptions, e.g., 'breakpoints' array lacks a description of what each element contains or what format is expected.
The 'continue' and 'step' tools mention 'fullContext' as an optimization hint, but the response schema (what fields are returned with vs without fullContext) is not documented. LLMs cannot reason about whether to set this flag without knowing what data they'll receive.
No idempotency guarantees documented. Are 'set-variable' or 'restart' idempotent? If an agent retries them on transient failure, will duplicate side effects occur?
The 'debug' tool supports two debuggers (delve, gdb) with different capabilities and mode support (e.g., 'GDB does not support source mode'). The schema doesn't gate 'mode' based on 'debugger' selection, so LLMs may try source mode with gdb despite the warning in the description.