Orinth Sidekick

A fully local coding agent I built so I could keep shipping when cloud AI was temporarily paused at work. One machine, one user, nothing sent anywhere.

Why I built it

My team briefly paused cloud AI while we reworked cost limits. Rather than lose momentum, I built a coding agent that runs entirely on my own machine, so I could keep working without sending anything to a hosted model or spending a cent per token. A constraint turned into a tool I still use.

What it is

The whole agent is a single stdlib-only Python file, about a thousand lines, with no dependencies to install. A small launcher script starts a 35B coding model (Ornith-1.0-35B, quantized) through llama.cpp's OpenAI-compatible server and auto-scales the context window to the machine's GPU limit. The agent talks to that local endpoint, reads and edits files, and runs entirely offline.

What I like about it

It is deliberately tiny and portable: no framework, no package manager, just Python's standard library and a local model. That makes it trivial to drop onto a new machine and trust, and it was a good exercise in how much a capable coding assistant really needs versus what a heavy toolchain adds.