A coding agent harness built for small models, not data centers

What it does

little-coder is a bundle of 20 extensions, 30 skill markdown files, and a Python benchmark harness layered on top of the minimal pi agent framework. It targets small local models served via llama.cpp, Ollama, or LM Studio, using standard read/write/edit/bash tools against your local project directory. Bash commands are pre-screened against a built-in safe-prefix whitelist before the underlying agent loop ever sees them.

The interesting bit

The harness treats scaffold–model fit as the critical variable: small-model-specific extensions automatically disable when you switch to a large cloud model, so they don’t get in the way. The authors claim a 9.7B Qwen running through this setup beat frontier entries on Aider Polyglot; the visible benchmark table shows it scored 45.56%, more than double the vanilla-Aider baseline of 19.11% for the same model.

Key highlights

• Ships as bundled extensions and skills atop pi, which remains a plain dependency rather than being forked
• Bash tool gated by a whitelist that excludes rm and sudo by default; extra prefixes can be added via environment variables
• Supports local inference stacks including llama.cpp (with MoE offloading), Ollama, and LM Studio, plus cloud providers
• Model and provider overrides live in user config files outside the package install path
• Includes a benchmark harness for reproducible evaluation on coding tasks

Caveats

• The documentation assumes prior familiarity with the pi framework’s extension and RPC model
• Even with a bun global install, the launcher still requires Node.js 22.19+ on your PATH
• Common build commands like make aren’t in the default bash whitelist and must be explicitly added via environment variables

Verdict

Worth a look if you want a coding agent that runs fully offline on modest hardware and you’re willing to tune a whitelist. Skip it if you already have a comfortable cloud-agent workflow and no interest in local inference.