JuliusBrussee/caveman-code

Frozen — terminal coding agent measured at 1.93× fewer tokens than Codex CLI. Still works; active development moved to JuliusBrussee/caveman (caveman wrap).

What it solves

Caveman Code is a terminal-based coding agent designed to drastically reduce LLM token usage and costs. It addresses the "waffle" (excessive filler text) produced by AI agents and the massive token overhead from tool outputs (like long git diffs or build logs) that are typically fed back into the model's context window.

How it works

The project employs a four-layer compression system to minimize token consumption:

  1. Caveman Mode: Forces the model to respond in terse, technical fragments without filler or hedging.
  2. Tool Budgets: Implements line caps for tool outputs (e.g., bash, grep, read) and strips ANSI colors and blank lines.
  3. Read Dedup: Fingerprints files per session so that repeated reads of the same file return a stub rather than the full content.
  4. RTK (Rust Token Killer): An optional external Rust binary that further compresses bash output before it enters the context.

Who it’s for

Developers who want a cost-effective, high-performance coding agent in their terminal that supports a wide range of LLM providers via API keys or OAuth (including Claude, ChatGPT, Gemini, and others).

Highlights

  • Autonomous Goal Loop: A "Ralph-style" autopilot that can work toward a goal autonomously with state tracking and budget caps.
  • Architect/Editor Split: Uses a slow, high-reasoning model for planning and a fast, cheap model for execution to reduce costs.
  • Plan Mode: A read-only mode where the agent explores the codebase and produces a plan without making edits.
  • Extensive Provider Support: Integration with over 20 providers and 6 OAuth flows for easy authentication.
  • Session Management: Supports session branching, forking, and git-like checkpoints for rolling back code and conversation history.
  • MCP Compatibility: Fully compatible with Model Context Protocol (MCP) servers and Claude Code's configuration formats.

Related

  • Project
    JuliusBrussee/caveman

    Caveman is a token‑saving wrapper for AI coding assistants. It provides a tiny “skill” that rewrites the assistant’s explanatory text in a terse style, and an optional local proxy that compresses large inputs (logs, JSON, diffs, etc.) before they reach the LLM provider. Benchmarks show up to 65 % fewer output tokens and ~33 % fewer input tokens, lowering cost and latency for agents like Claude Code, Codex, Gemini, and many others. Installation is a one‑liner for the skill and a global npm install for the proxy; the tool works with 30+ agents without altering their config files.

  • Project
    headroomlabs-ai/headroom

    A local context compression layer for AI agents that reduces prompt and response tokens for JSON, code, and text while preserving critical information.

  • Project
    JuliusBrussee/cavekit

    A token-efficient, spec-driven development framework for Claude Code that uses a compressed encoding and a single durable spec file to maintain project memory.

  • Project
    jfrog/boost

    Boost is a command wrapper that compresses noisy shell output into structured context for AI coding agents, reducing token costs without sacrificing task performance.