Little-Coder NEW

Little-coder is a Pi-based coding agent for running smaller local LLMs via llama.cpp or Ollama.It provides a Python/Node.js CLI and TypeScript extensions that adapt cloud-style coding agent workflows for 5–25 GB models served locally.

Core mechanisms preserved from the whitepaper include the write-vs-edit invariant, per-turn skill injection, algorithm-cheat-sheet injection, thinking-budget cap, output parsers, quality monitor, per-model profiles, and evidence-aware compaction.

The repository contains a Pi port, build and serve instructions for llama.cpp (GPU and MoE options), Ollama setup, model fetching steps, and reproducible tags for reported benchmark runs.Developers and researchers can use little-coder for local code generation, on-device development, model benchmarking, and reproducing published results with included configs and documentation.

Benchmarks and whitepaper documents support evaluation across local LLMs such as Qwen variants and custom GGUF models.Quick-start steps and scripts facilitate setup on laptops, GPUs, and Raspberry Pi devices for integrated developer workflows and edge deployments.

Key Features

• Runs local small LLMs via llama.cpp or Ollama (Pi-based coding agent)
• Python and Node.js CLIs plus TypeScript extensions for local cloud-style coding agent workflows
• Core agent mechanisms: write-vs-edit invariant, per-turn skill injection, algorithm-cheat-sheet injection, thinking-budget cap, output parsers, quality monitor, per-model profiles, evidence-aware compaction
• Build/serve instructions and scripts for llama.cpp (GPU and MoE) and Ollama, plus model fetching and reproducible benchmark tags
• Benchmarking and evaluation support for local LLMs (e.g., Qwen variants and custom GGUF models) with reproducible configs and documentation

Use Cases

• Build an offline on-device coding assistant on a Raspberry Pi using little-coder's Python/Node CLIs and TypeScript extensions to generate, run, and debug code locally with 5–25 GB LLMs (llama.cpp or Ollama) while using evidence-aware compaction for more reliable suggestions
• Run reproducible benchmarks to compare low-memory LLMs and optimize edge deployments using little-coder's benchmark suite and build/serve guides, producing repeatable performance reports to pick the best model for constrained hardware
• Automate secure local code generation and evaluation in CI/CD by integrating little-coder's CLI developer workflows and on-device evaluation tools to produce audited, reproducible artifacts for edge LLM deployment without sending code to the cloud

Who is it for?

• Local llm developers
• Embedded developers
• Benchmark engineers
• Reproducibility researchers
• Students