little-coder
A coding agent tuned for small local models, with a minimal system prompt, a planning mode and background sub-agents for research
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- Runs bash commands and edits project files on the agent's behalf
- Works with an arbitrary local or cloud model whose behavior the user must watch themselves
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/little-coderDetects the agents on your machine, checks the risk and pins the version.
Other ways to install
Install the tool
npm install -g little-coderInstall: npm install -g little-coder (or curl -fsSL https://raw.githubusercontent.com/itayinbarr/little-coder/main/install.sh | bash). Run it in a project: little-coder, or name a model explicitly: little-coder --model llamacpp/qwen3.6-35b-a3b.
Other ways from the author
npm install -g little-coderRequires Node.js 22.19 or higher.
This is third-party code. Review the repository files before installing.
What it does
little-coder is a standalone terminal coding agent built on the minimal pi engine: four core tools (read, write, edit, bash) and a short system prompt that keeps the cold-start context compact and behavior predictable. On top of pi it adds extensions, skills and a benchmark harness aimed at making small models like Qwen 3.6 35B running through llama.cpp perform close to top-tier models on coding tasks. It has a planning mode that first researches a task with sub-agents and asks clarifying questions before touching code, and a deep-research mode that produces a separate cited report on an external topic.
Who it is for. For developers who want a coding agent for local or cheap models instead of constantly relying on expensive cloud LLMs.
Good fit when
- You need a coding agent that works well with small local models via llama.cpp or Ollama
- You want a planning mode with clarifying questions before the agent starts changing code
- You need a separate coding agent for quick tasks without spending your main cloud agent's quota
Not a fit when
- You need a graphical interface rather than a terminal agent
- Node.js below 22.19 and updating the runtime is not an option
- You already have an established workflow on Claude Code, Cursor or Codex and do not need another standalone agent
Example request
Launch little-coder on the local qwen3.6-35b-a3b model and use plan mode to figure out what this module doesLimitations
Requires Node.js 22.19 or higher even when installed via bun, because the launcher itself is a Node script. It is canonically tuned for a local llama.cpp server running Qwen3.6-35B-A3B, which must be set up separately, though cloud models like Anthropic and OpenAI are also supported. Extensions installed globally via pi install are not picked up by default, because little-coder launches pi with the --no-extensions flag.
How to disable. Remove the global package: npm uninstall -g little-coder (or bun remove -g little-coder), and delete the ~/.config/little-coder folder if needed.
Security check
- Runs bash commands and edits project files on the agent's behalf
- Works with an arbitrary local or cloud model whose behavior the user must watch themselves
README in short
The README describes little-coder as a coding agent tuned for small local models, built on top of the minimal pi engine. It installs with one command via npm, bun or an install script, with no separate project build. It details a planning mode with sub-agents and clarifying questions before code changes, a deep-research mode producing a cited report, a status line tracking token usage, and situational tool-skill cards the agent surfaces. Extensions default to a bundled set for predictable behavior and a compact starting context. The author links a separate write-up with Aider Polyglot benchmark results. Apache-2.0 licensed.
FAQ
Is a local model required?
No, little-coder also supports cloud models through the --model flag, such as anthropic/claude-haiku-4-5 or openai/gpt-4o-mini, but it is tuned and tested primarily for a local llama.cpp server running Qwen3.6-35B-A3B.
How does little-coder differ from pi, which it is built on?
pi is the minimal agent engine: the loop, the TUI, a multi-provider API. little-coder is pi plus its own extensions, skill files and a benchmark harness tuned to make small models behave predictably with a small context.
Related
Open-source personal AI assistant on your own machine: answers in Telegram, Slack, Discord and WhatsApp, extended with skills and plugins
A self-improving agent from Nous Research with a TUI, messaging gateway, cron jobs and skills it writes itself
An open source coding agent for the terminal and desktop with build and plan modes
Open prompt library with a Claude Code plugin, MCP server and CLI: search, fetch and improve prompts and skills from an agent