itayinbarr/little-coder

What it solves

little-coder is a coding agent designed specifically for small, local LLMs. It addresses the gap where frontier models (like GPT-4) are typically required for complex coding tasks, allowing users to run a high-performance coding assistant directly on their laptop using models like Qwen3.6-35B-A3B.

How it works

Built on top of the pi agent framework, little-coder adds a specialized set of 30+ extensions, 30 skill markdown files, and a Python benchmark harness. It uses a launcher that ensures a predictable, minimal context (around 7k tokens) by loading only the bundled extensions.

Key operational mechanisms include:

• Plan Mode: A research-first approach where the agent researches a task, asks clarifying questions, and creates a plan before implementation.
• Sub-coders (dispatch): The ability to spawn isolated, read-only child sessions to research questions without cluttering the main conversation.
• Deep Research: A dedicated mode for external and online research that produces a cited markdown report.
• Tool Skill Cards: Dynamic injection of short usage guides for tools based on current needs (error recovery, recency, or intent).
• Background Jobs: Long-running commands (like builds or training) run in the background, waking the model only on specific events.

Who it’s for

Developers who want a powerful coding agent that runs locally on their hardware (via llama.cpp, Ollama, or LM Studio) to maintain privacy, reduce costs, or experiment with small-model performance.

Highlights

• Local-First Optimization: Tuned for models like Qwen3.6-35B-A3B, with specific configurations for MoE (Mixture of Experts) to fit large models into limited VRAM.
• Flexible Model Routing: Supports local providers (llama.cpp, Ollama, LM Studio) and cloud providers (Anthropic, OpenAI) with easy override configurations.
• Hiearchical Task Management: Separates planning from implementation to preserve context window efficiency.
• Advanced TUI: Includes a detailed status line tracking cumulative tokens, cache-hit rates, and context window usage.
• Extensible: Users can add their own extensions via a config directory.