JuliusBrussee/caveman

🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman

What it solves

Caveman reduces the number of tokens used by AI agents, lowering costs and increasing speed. It targets both the input tokens (what the agent reads) and output tokens (what the agent says), preventing agents from being overly verbose or processing redundant data.

How it works

Caveman employs several mechanisms to shrink token usage:

  • Caveman Proxy: A local proxy that intercepts traffic between the agent and the AI provider. It uses a compression engine to shrink payloads (JSON, logs, code, diffs, and text) by removing noise while keeping essential information. Original bytes are stored locally for byte-exact recovery if needed.
  • The Skill: A prompt-based extension that instructs agents to answer in a terse, "caveman-speak" style, significantly reducing output tokens without losing technical accuracy.
  • Pixel Mode: Converts text-based skill prompts into PNG images. Since models can read images, this reduces the input token tax paid when loading large skill prompts.
  • Caveman Learn: Scans local agent history to identify "token sinks" (inefficient patterns) and provides a plan to fix them via an agent-driven implementation process.
  • MCP Server: Provides tools like caveman_compress and caveman_retrieve to any MCP-compatible host.

Who it’s for

Developers and power users of AI coding agents (such as Claude Code, Codex, Gemini CLI, Aider, and others) who want to reduce their API spend and improve the responsiveness of their agents.

Highlights

  • Input Compression: Reduces provider-reported input tokens by up to 33.2% in specific benchmarks.
  • Wide Compatibility: Natively wraps 8 popular agents and supports 30+ others via a skill.
  • Output Reduction: Achieves significant savings on output tokens (averaging 65% in provided examples) by enforcing terseness.
  • Byte-Exact Recovery: Ensures that compressed data can be restored to its original form without loss of information.
  • Local-First Privacy: Compression and history scanning happen locally; no account is required for the core functionality.

Related

  • Project
    JuliusBrussee/caveman-code

    A terminal coding agent that uses a four-layer compression system to drastically reduce LLM token usage and costs while maintaining performance.

  • Project
    JuliusBrussee/cavemem

    A cross-agent persistent memory system for coding assistants that uses compressed local storage and MCP to allow AI agents to remember information across sessions.

  • Project
    headroomlabs-ai/headroom

    A context compression layer for AI agents that reduces input and output tokens by 15-95% using content-aware compressors and a reversible caching system.

  • Project
    alexgreensh/token-optimizer

    Token Optimizer is a zero‑dependency plugin for AI coding assistants (Claude Code, Copilot, Codex, etc.) that automatically compresses command output, config files, and model responses, checkpoints before the assistant’s own context compaction, and restores saved context. It tracks token usage, gives a live dashboard of cost and quality scores, and provides slash‑commands for audits and fixes—all with no telemetry and cross‑platform support.