I build agent harnesses, inference fallbacks, and small tools people actually run. Most of the work is making models hold up in real workflows: long-running loops, MCP and tool interop, and routing that fails over instead of dying on a bad upstream.
Cesium — Cursor-inspired open-source agent (TypeScript)
An OSS agent surface with multiple harnesses (native, Cursor, OpenCode, Codex, and others). Self-host it, keep work running while you are away, and hook it into GitHub and Linear. This is where most of my time goes.
Model Proxy — any model, anywhere, with fallbacks (TypeScript, Bun)
I wanted Claude Code and similar clients to talk to whatever model I actually use. Model Proxy translates OpenAI and Anthropic APIs, streams cleanly, keeps tool-calling and multimodal payloads intact, and fails over at the API key, provider, and model layers.
Codex Meter — Android Codex usage monitor (Java)
A native Android client and widget suite for Codex allowance: rolling windows, reset times, credits, and optional notifications. Built so I could see usage on my phone without opening a browser.
- MCP-Base — TypeScript MCP server template with OpenAI and Anthropic-style methods, plus streamable HTTP and local transport
- Insta-AI-Correct — AutoHotkey shortcut that Groq-corrects the text you just wrote
- Noetic — personal WIP: local-first notes, tasks, and calendar with a first-party agent. I use it daily; the public repo is not up yet
- Languages: TypeScript, JavaScript, Python, Java, CSS
- Runtime / frameworks: Bun, Node.js, Next.js, React, FastAPI
- Platforms: GitHub, Docker, Vercel, Railway, Android