eloquent-cavespeak

Credit: forked from JuliusBrussee/caveman. Julius built the compression concept. This project adds the rehydration layer.

Claude speaks cavespeak. Gemma rehydrates. You read prose. Token cost stays low.

---

How it works

Claude Code CLI
    │  ANTHROPIC_BASE_URL=http://localhost:3000
    ↓
proxy.js (Node, local)
    ├─ injects cavespeak system prompt → Anthropic API
    ├─ tool_use responses → PASS THROUGH (no rehydration)
    └─ end_turn responses → Ollama (Gemma 4) → polished prose → CLI

Claude compresses internally. Gemma translates back to English. Tool calls — file reads, bash commands, searches — pass through untouched so the pipeline stays fast.

---

Two modes

/skill cavespeak — compression only. Responses are terse caveman-speak. Cheap and fast, but raw.

/skill eloquent-cavespeak — full pipeline. Same token savings, polished output. Requires Ollama running locally.

---

Setup

Prerequisites: Node.js 18+, Ollama

# 1. Install Ollama
brew install ollama   # or: https://ollama.com/download

# 2. Install the skills into Claude Code
npx skills add coderlevelup/eloquent-cavespeak

# 3. Clone and install dependencies
git clone https://github.com/coderlevelup/eloquent-cavespeak
cd eloquent-cavespeak
npm install

Then in Claude Code, run the skill — it handles starting the proxy, checking Ollama, and pulling the model automatically:

/skill eloquent-cavespeak

Follow the prompt it returns to restart Claude Code with ANTHROPIC_BASE_URL=http://localhost:3000.

---

Environment variables

Variable	Default	Description
PORT	3000	Proxy listen port
OLLAMA_HOST	http://localhost:11434	Ollama base URL
GEMMA_MODEL	gemma4:latest	Model used for rehydration

---

The meta moment

Run /skill eloquent-cavespeak inside this repo, then ask Claude to explain how the proxy works. It will explain itself — compressed upstream, polished on the way out.