smallcode

smallcode is a terminal coding agent built for small, local language models, the 8B to 35B range you can run on your own hardware, not frontier models like Claude or GPT. Most agents assume a huge, reliable model behind them. smallcode assumes the opposite and engineers around it. MIT-licensed and free.

The adaptations are practical: a managed context budget instead of dumping everything at the model, forgiving parsing for messy tool calls, TODO-file planning, and search-and-replace patches instead of full-file rewrites. It's actively maintained, with real releases and a benchmark harness. The project claims 87% on a benchmark with a 4B-active model, though it doesn't spell out the suite or baseline, so take the number as a signal, not proof.

The catch: a coding agent is only as good as the model driving it, and small local models still make mistakes a frontier model wouldn't. If you want to keep code on your own machine or avoid API bills, this is a serious attempt at making that work. If you just want the best results, a frontier model with a mature agent like aider will still beat it.