juliusbrussee/caveman
JuliusBrussee/caveman is an open-source proxy and skill designed for AI coding agents to reduce token usage and operational costs while aiming to preserve technical accuracy. The tool optimizes communication by enforcing concise output formatting and compressing ingested content such as logs, JSON payloads, and test outputs before processing. Its core feature set includes automated model selection based on task difficulty, code script persistence through block execution, web page compression, and prompt cache maintenance. The GitHub snapshot shows the repository is not archived and records an Apache-2.0 license.
Targeted at software engineers and teams utilizing tools like Claude Code, Codex, and Gemini CLI, the README states that code syntax, paths, and critical instructions are preserved; verify important output because this is not a guarantee. Verified prerequisites include Node.js version 22.13 or higher across macOS, Linux, and Windows environments. Users can configure the tool using the command `npx @caveman-ai/cli@latest` or install the individual skill via `npx skills add JuliusBrussee/caveman -g`. The documentation evaluates this approach against simple prompting techniques like requesting concise answers, and reports token savings in its own tests; results can differ by task. RTK is another token-optimization tool with a different scope.
Data as of : 117005 stars · 6799 forks.