Caveman Core

MCP server that teaches LLMs to respond in compressed caveman-style prose.
Cuts output tokens ~65% while keeping full technical accuracy. Zero external dependencies.

What it exposes

Prompts

Name	Effect
caveman	Full mode — drops articles, fillers, fragments OK
caveman_lite	Lite mode — removes filler/hedging, keeps articles and full sentences
caveman_ultra	Ultra mode — maximum compression, one word when one word enough

Resources

URI	Content
caveman://rules	Full ruleset text
caveman://rules/lite	Lite ruleset text
caveman://rules/ultra	Ultra ruleset text

Tools

Name	Description
compress_text	Post-process any text through the compression algorithm

Install

{
  "mcpServers": {
    "caveman": {
      "command": "node",
      "args": ["/path/to/caveman-core-mcp/src/index.js"]
    }
  }
}

Or with npx after publishing:

{
  "mcpServers": {
    "caveman": {
      "command": "npx",
      "args": ["caveman-core-mcp"]
    }
  }
}

Usage

Inject via prompt (recommended)

Use the caveman prompt at session start. The host sends the ruleset as context and the LLM adopts compressed mode for the rest of the session.

Inject via resource

Read caveman://rules and include the text in your system prompt manually.

Post-process text

Call the compress_text tool to strip filler from any already-generated text.

{
  "name": "compress_text",
  "arguments": {
    "text": "Sure! I'd be happy to help you with that. The issue you're experiencing is likely caused by a missing semicolon."
  }
}

Returns:

{
  "content": [{ "type": "text", "text": "Issue caused by missing semicolon." }],
  "_meta": { "before": 95, "after": 35, "savedPercent": 63 }
}

Requirements

• Node.js >= 18
• No npm dependencies

Origin

Based on caveman by Julius Brussee (MIT).

• src/compress.js — derived from src/mcp-servers/caveman-shrink/compress.js
• src/skill.js — derived from skills/caveman/SKILL.md
• src/index.js — new standalone MCP server built on top of the above