Caveman is a toolset designed to reduce token consumption and costs in AI coding agents by compressing both agent outputs and inputs. It operates through three primary mechanisms: a "skill" that forces agents to use terse, "caveman-style" prose; a local proxy that shrinks incoming data (logs, JSON, diffs, and web pages) before it reaches the model; and middleware for developers to compress tool results within their own applications. Technical evaluations show significant efficiency gains: Adobe Research reported that caveman-style output can reduce costs by 1.4× to 3×, while JetBrains testing of the skill alone found a 10% cost reduction with no measurable impact on quality. A comprehensive 54-run benchmark using Claude Code demonstrated that the proxy mechanism achieved a 33.2% reduction in total input tokens across various workloads, including a 55.1% reduction in a CSV outlier hunt. The system maintains data integrity by storing original, uncompressed payloads in a local SQLite database, allowing the agent to retrieve the full context via a recovery handle when necessary. It supports over 30 agents, including Claude Code, Cursor, and Gemini, and offers a "pixel mode" that converts skills into PNG images to further reduce token overhead by approximately 61%.