Low-level infrastructure for AI agents — debuggers, pipelines, and the plumbing between them.

I build tools that give agents real capabilities against real systems: a live WinDbg engine over MCP, an autonomous plan→execute→judge loop. Written in Rust and TypeScript, MIT-licensed, shipped.

Selected work

02 / open source

windbg-mcp

→

An MCP server exposing WinDbg/DbgEng to AI agents over stdio — driving a live debugger engine for user-mode, kernel, crash-dump, and Time Travel Debugging workflows.

30+ tools single-thread engine TTD replay crash-dump triage

Architecture

AI agent —stdio→ MCP server
rmcp tools —mpsc→ engine thread
DbgEng → target
user / kernel / dump / TTD

Rust
Windows x64
MIT · v0.1.3

plan-execute-judge

→

A generic plan→execute→judge pipeline for the Claude Agent SDK or Codex SDK. Fresh agent runs with real control flow — only a brief, a plan, and a typed verdict cross phase boundaries, never a transcript.

multi-agent fan-out typed verdict baseline diff Claude or Codex

Pipeline

task → research → plan → execute → judge ↶ fail loops back

TypeScript
Node ≥ 22.9
MIT

win-kexp — the low-level DbgEng bindings, and more

soon

About

I work close to the metal — debugger engines, agent SDKs, and the interfaces that let language models drive them. Everything here is open source and built to be read.

github.com/glslang ↗