Skip to content

Prior Art

Several projects bridge MCP to DAP. All solve the plumbing problem. None address agent ergonomics — which is Krometrail's primary focus.

Projects

ProjectLanguageDebuggersKey contribution
mcp-debuggerTypeScriptPython, JS, Rust, GoDynamic adapter loading, clean architecture
mcp-dap-serverGoGo/DelveAuto context return, capability-gated tools
AIDBPythonPython, JS, JavaFramework auto-detection, launch.json reuse
debugger-mcpRust5 languagesReal agent integration tests (Claude Code, Codex)
dap-mcpPythondebugpy, lldbConfig-driven adapter selection

What Each Got Right

mcp-dap-server's automatic context return. After every step or continue, it calls getFullContext — stack, scopes, and variables in one response. No separate tool calls required. Krometrail follows this pattern: every execution control tool returns the viewport automatically.

mcp-debugger's get_local_variables. Collapses the standard DAP dance (stack → scope → variables) into one call. Also filters out noisy __builtins__/__proto__ entries. Krometrail's viewport does this automatically on every stop.

AIDB's framework auto-detection. Auto-identifies pytest, jest, django, spring so agents don't need to configure debug adapters for common cases. Krometrail implements the same pattern.

debugger-mcp's agent integration tests. Tests run against real Claude Code and Codex agents (not mocked clients). 10 test matrices across 5 languages. Proves the approach works end-to-end. Krometrail's agent harness follows this model.

mcp-dap-server's capability-gated tools. Tools only appear when the underlying debugger supports them (restart, set-variable, disassemble). Agents never see tools that won't work.

What No One Solved

The token problem. Every existing project returns raw DAP state with no token awareness. mcp-dap-server's getFullContext dumps all scopes and all variables with no truncation — a moderately complex program produces thousands of tokens per stop.

Session intelligence. No project maintains an investigation log, implements viewport diffing, or progressively compresses output over long sessions.

A CLI interface. All existing projects require MCP server configuration. The Krometrail CLI (session daemon, full command parity with MCP) is unique and addresses environments where MCP setup is inconvenient.

Browser observation. No existing MCP-DAP project captures browser activity. Krometrail's CDP recording, session investigation tools, and framework state observation are entirely novel in this space.

Key Lessons Applied

  1. Automatic context return after every execution control operation
  2. Hide DAP internals — agents should not need to know variablesReference, frameId, or scope hierarchies
  3. Source context must be automatic, not a separate tool call
  4. Minimal tool surface is sufficient — debugger-mcp proves 7 tools can work
  5. Framework detection reduces friction for common test frameworks
  6. Keep adapters thin — the proxy complexity in mcp-debugger isn't necessary

Last updated: