Summary

Caveman Code tackles the high token consumption and costs associated with LLM coding agents. It offers a ~2x more token-efficient alternative to tools like Codex, enabling significant cost savings and faster developer interactions. The project targets engineers and power users optimizing LLM-powered development workflows.

How It Works

The core innovation is a multi-layered "Caveman Mode" compression strategy, drastically reducing token counts for model responses and tool outputs. Techniques include terse technical fragments, per-tool line caps, ANSI stripping, and read deduplication. An optional Rust binary (RTK) further compresses shell output. This minimizes context window usage and API costs. An architect/editor model split enhances cost efficiency.

Quick Start & Requirements

Install via npm: npm install -g @juliusbrussee/caveman-code. Requires LLM provider API keys (Anthropic, OpenAI, etc.) or OAuth for services like Claude Pro/ChatGPT Plus. Docker installation is supported. See docs/getting-started/installation.md for details.

Highlighted Details

• Token Efficiency: Achieves ~1.93x fewer tokens than Codex CLI, with benchmarks showing up to 94% tool output reduction and potential savings over 1 million tokens per session.
• Advanced Features: Includes autonomous goal loop (--goal), read-only plan mode (--plan), parallel subagents, session branching with rollback, and persistent semantic memory (cavemem).
• Broad Provider Support: Integrates with 20+ LLM providers via API keys or OAuth, including major platforms and custom endpoints.
• MCP Compatibility: Adopts and extends Claude Code's MCP formats, simplifying migration.

Maintenance & Community

Maintained primarily by Julius Brussee, it's a significant fork of pi-code (Mario Zechner). Acknowledges inspiration from Aider and Claude Code. Community interaction via GitHub Issues and Releases.

Licensing & Compatibility

Distributed under the permissive MIT License, allowing for commercial use and integration into closed-source projects. MCP format compatibility aids migration from Claude Code.

Limitations & Caveats

Relies on external LLM provider APIs, requiring key/subscription management. Effectiveness of some compression layers may vary. Development appears heavily centered around its primary author.