itayinbarr/little-coder

What it solves

little-coder is a coding agent designed specifically for small local models, ensuring they can perform complex coding tasks without the overhead of massive frontier models. It addresses the problem of "scaffold-model fit," where the agent's framework (the scaffold) is too heavy for the smaller models to reliably follow instructions or manage context windows.

How it works

Built on top of the pi agent substrate, little-coder adds a layer of specialized extensions, skill markdown files, and a benchmark harness. It uses a curated set of bundled extensions to keep the cold-start context small (around 7k tokens) and behavior predictable.

Key mechanisms include:

• Plan Mode: A separate research and planning phase where the agent researches the request, asks clarifying questions, and writes a plan before implementation. This prevents research noise from cluttering the implementation context.
• Sub-coders (dispatch): The ability to spawn isolated child sessions to research questions without cluttering the main conversation.
• Tool Skill Cards: Dynamic injection of short usage guides for tools based on current needs (error-recovery, recency, or intent).
• Local Model Integration: Native support for llama.cpp, Ollama, and LM Studio, with specific optimizations for MoE (Mixture of Experts) models like Qwen3.6-35B-A3B.

Who it’s for

Developers who want a powerful coding agent that runs locally on their own hardware (e.g., using llama.cpp or Ollama) and can handle complex tasks using smaller, more efficient models.

Highlights

• Local-First: Optimized for local LLMs, specifically tuned for Qwen models.
• Coded-Plan-Implement: A workflow that separates research/planning from execution to save context window space.
• Deep Research: A dedicated mode for external and online research that produces a cited markdown report.
• ** uma-low overhead**: Minimal cold-start context (7k tokens) to keep small models focused.
• Background Jobs: Supports long-running commands (builds, servers) that run in the background and wake the model on events.
• Extensible: Users can add their own extensions in a config directory.