little-coder

Why we included this project

This coding agent is tuned to run well on the hardware most developers actually have, not the cluster a frontier model would demand. It layers roughly thirty extensions and thirty skill markdown files onto the pi substrate, and the launcher keeps cold-start context around 7k tokens so a small local model stays productive instead of fighting prompt noise. The authors back the design with public numbers: an 8GB machine running Qwen3.6-35B-A3B finished above Gemini CLI and Gemini 2.5 Pro on the official terminal-bench 2.0 leaderboard. Small teams get an offline, no-API companion whose behavior stays predictable, since exactly the shipped extension set loads and nothing in the working directory changes the setup mid-task.