gemma-coder

Save your frontier tokens for thinking. A free local model writes every line of code — and tight specs make open models write it well.

gemma-coder is an Agent Skill — one folder that plugs into Claude Code, Google Antigravity (CLI & IDE), OpenAI Codex CLI, and any other agent that supports the open SKILL.md standard. It’s built on two ideas:

1. Save tokens. Most of what a coding agent burns tokens on is typing code out, not thinking. Here the frontier model spends tokens only on what it’s uniquely good at — architecture, per-file specs, code review, tests — while a free local model (Gemma, Qwen, anything you can run) writes every line of application code on your own machine. Zero API cost per generated line.
2. Boost open-model code quality. Small models don’t fail on well-specified files; they fail on vague prompts. The skill enforces a strict contract — the agent never writes application source itself, and when generated code fails tests it improves the spec and re-delegates — so a 12–27B local model reliably produces code it could never write from a one-line ask.

The only requirement besides an agent: Python 3.9+ (standard library only) and one local model runtime (step 2 below — ollama is the easiest).

---

Step 1 — Install the skill

Quickest: npx (no clone)

npx skills add vibecoding-skills/gemma-vibecoding-skills

Installs the skill into your agents’ skill directories via the skills CLI. Update later with npx skills update (npx installs are snapshot copies).

Or from a clone (adds the gemma-coder command to your PATH):

git clone https://github.com/vibecoding-skills/gemma-vibecoding-skills
cd gemma-vibecoding-skills
./install.sh                # symlinks into every agent skill dir found on your machine

install.sh puts a gemma-coder command on ~/.local/bin (make sure that’s on your PATH). Everything is symlinked to the checkout, so a plain git pull updates every agent at once — no reinstall needed. Run ./install.sh --copy instead for a standalone copy that survives deleting the clone (update it by pulling and re-running ./install.sh --copy).

Step 2 — Set up local Gemma

1. Install ollama — download from ollama.com/download (macOS/Windows installer or curl -fsSL https://ollama.com/install.sh | sh on Linux). The app runs the server automatically; otherwise start it with ollama serve.

2. Pull a Gemma model sized to your RAM:

   ollama pull gemma4:12b-nvfp4   # 7.7 GB — fits comfortably on 16 GB machines
   ollama pull gemma4:26b-nvfp4   # ~17 GB free RAM — best quality/speed balance
   

   (Full device guide: Recommended Gemma by device.)

3. Point the skill at it:

   gemma-coder setup            # clone install
   # npx install (no command on PATH):