rig
A small, single-binary agent runtime that points at any OpenAI-compatible model, local or hosted, reads and edits a repository's code, runs its tests and remembers what it learned, with a loop implemented in a few hundred lines.
rig is a small agent runtime distributed as one binary with no separate runtime to install, meant to turn a model the user already runs, local or hosted, into an agent that can read a repository's code, edit it, run its tests, and remember what it learned across turns. It connects to anything that speaks the OpenAI-compatible API, including llama.cpp, llama-swap, Ollama and OpenRouter, configured through a small set of local JSON files that set which model is active, each model's context window and token budget, and how many background workers run on which model.
The project describes its whole agent loop, batching included, as a few hundred lines of code, deliberately kept small enough to read before trusting it, with tools, guardrails, schedules and background workers implemented as separate pieces that plug into that loop. Tool calls run concurrently, and the project states that most of the tokens it serves in real use come from cache rather than being generated fresh. It installs with a shell script and can update itself in place afterward, and a session can be closed on one device and picked up again from another.