A Debug Adapter Protocol (DAP) enabled Model Context Protocol (MCP) server that bridges debugging tools with AI systems, providing a standardized context for LLMs.
dap-mcp provides 14 debugging tools with clear descriptions and basic parameter schemas. However, several critical definition quality issues prevent a higher score: (1) Most tool parameters lack descriptions, only 3 of 14 tools have fully documented parameters (set_breakpoint, remove_breakpoint, view_file_around_line, evaluate, change_frame). Tools like launch(), continue_execution(), step_in(), step_out(), next(), terminate() have empty parameter schemas with no description text. (2) Output schemas are not documented anywhere, the rubric requires documentation of what fields to expect in responses, but the source shows only `.render()` calls with no schema definitions. (3) Four tools (get_launch_config, launch, remove_all_breakpoints, list_all_breakpoints) have empty input schemas `{"type":"object","properties":{}}` but lack parameter descriptions. (4) Tool descriptions are generally adequate (50-150 chars), meeting the 10-1024 char baseline, but lack depth on WHEN to use the tool vs alternatives and what to do with responses. (5) Error handling is present (FunctionCallError class) but sparse, most tools lack explicit recovery guidance. The server does follow good naming conventions (verb_noun like set_breakpoint, remove_breakpoint, continue_execution) and demonstrates some composition patterns (breakpoint lifecycle tools). However, missing output documentation and inconsistent parameter descriptions pull the overall score to the 'poor' range.
Change the active stack frame.
Continue execution of the debugged program.
Evaluate an expression in the current debug context.
Returns the user provided launch configuration along with its detailed schema for a DAP-compatible debugger. The schema includes descriptions for each field.
Launch the debuggee program. Set breakpoints before launching if necessary.
List all breakpoints.
Step to the next line.
Nine tools have empty or missing input schemas (`{"type":"object","properties":{}}`). Affected: launch, remove_all_breakpoints, list_all_breakpoints, continue_execution, step_in, step_out, next, terminate.
No output schemas documented anywhere. Rubric requires tools to document what fields are returned so LLMs can chain calls and extract data. The code shows `.render()` calls (XML rendering) but no schema definitions for response structure. LLMs cannot plan follow-up calls without knowing what data they receive.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 53 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 28 | - | v1 |
Remove all breakpoints.
Remove a breakpoint at the specified file and line.
Set a breakpoint at the specified file and line with an optional condition.
Step into the next function call.
Step out of the current function.
Terminate the debugging session.
View source code around a specified line.
Tools without parameters (launch, remove_all_breakpoints, list_all_breakpoints, continue_execution, step_in, step_out, next, terminate) lack descriptions explaining WHEN to use them instead of similar tools and what happens when called. Most of these are under 20 chars (e.g. 'Continue execution of the debugged program' is 40 chars).
Error handling responses (FunctionCallError) are minimal. No recovery guidance, when file not found, response is just 'File (path) not found'. Example fix: 'File (path) not found. Ensure the path is relative to sourceDirs or provide an absolute path.'
The 'terminate' tool is destructive but lacks confirmation logic or dry-run. No confirm_before_execute pattern present.
Output format is XML strings (via render_xml()) rather than structured JSON objects. Rubric pattern:response-shaper requires structured objects with typed fields. XML responses require LLMs to parse unstructured text, wasting tokens and increasing hallucination risk.