MCP Server exposing DAP debugging tools. Lets any MCP client debug programs (Python, Node.js, Java, browser JS) like a developer in VS Code: set breakpoints, step through code, inspect variables, evaluate expressions.
Server has 16 tools with complete input schemas and descriptions. Naming follows verb_noun pattern (debug_*). Descriptions are detailed but verbose (200-400+ chars, exceeding the 10-200 char baseline for LLM efficiency). All parameters have types and descriptions. However, output schemas are not formally documented in the code, only described in prose. Error handling is minimal; tools return JSON but lack recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Composition is sound (each tool does one thing), but descriptions are overly prescriptive with mandatory formatting templates that bloat the prompt.
Clear all breakpoints or breakpoints at specific lines.
Resume execution until the next breakpoint or program termination.
Evaluate an expression in the current execution context.
Expand a variable to see its nested properties or array elements.
Retrieve captured program output (stdout, stderr) since the last call.
Launch a program under the debugger. Args: program: Path to the source file to debug (.py, .js, .ts, .java), or a URL (http://...) for browser debugging. args: Command-line arguments for the program. cwd: Working directory (defaults to script's directory). For browser mode, used as webRoot for source maps. stop_on_entry: Pause at the first executable line (default: True). port: DAP port (default: 5679). language: Debug language: 'python', 'node', 'browser', 'java'. Auto-detected from extension/URL if omitted. python_path: Path to Python interpreter (Python only). Auto-detects project .venv if omitted. browser_path: Path to Chrome/Chromium (browser only). Auto-detects if omitted. headless: Run browser in headless mode (browser only, default: False). java_home: Path to JDK home (Java only). Auto-detects from JAVA_HOME or PATH if omitted. Returns: JSON: {host, port, program, pid, language-specific keys (python/java/adapter), stopped_at?, reason?} or {error}.
Output schemas not formally documented. Tool descriptions mention return structure in prose (e.g., 'Returns: JSON: {host, port, program, pid, ...}') but no structured schema definition is visible in code. LLMs cannot reliably parse prose schemas to extract fields for downstream tool chaining.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools are marked with Risk levels (WRITE, READ_ONLY) in metadata but not exposed via MCP tool annotations. Agents cannot distinguish safe read-only tools from destructive ones without parsing risk labels.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Reset the debug session (disconnect and clear all state). Use before launching a new program.
Set breakpoints at specific lines in a file.
Get the current call stack (all frames from the current execution point to the entry point).
Get the current debug session status (running, stopped, terminated, etc.).
Step into the next function call.
Step out of the current function (resume until the function returns).
Step over the next line (execute the current line and stop at the next).
Stop the current debug session immediately (SIGTERM) and return program output. Returns: JSON: {stopped: true, output: string|null}.
Inspect variables in the current scope (locals, globals, or a specific scope).
Wait for a debug event (breakpoint, step, exception, output, etc.) with a timeout.
Descriptions are overly verbose (200-400+ chars) and include mandatory formatting templates. Baseline for LLM-optimized descriptions is 50-200 chars. The embedded display templates (_RICH_STOP, _RICH_VARS, etc.) bloat descriptions and waste tokens. Move formatting guidance to server instructions, not tool descriptions.
Error handling lacks recovery guidance. Tools return JSON but do not include actionable error messages. E.g., if debug_launch fails due to missing Python interpreter, the response should suggest 'Set python_path parameter or ensure Python is in PATH.'
Parameter 'language' in debug_launch accepts free-form string with enum hint in description only. Should be declared as enum: ['python', 'node', 'browser', 'java'] in schema to prevent hallucinated values.