Caveman: Recoverable Context Compression for Coding Agents

What

Caveman is Caveman is an open-source skill and local runtime that adds recoverable context compression and token usage measurement to coding agents. It targets engineering teams, coding agent users, and AI developers, and supports Claude Code plus 30+ agents.

Core capabilities

• Output compression: output is written compressed, reducing large logs, files, and tool results that travel into later turns.
• Tool budgets: a cap per tool, per turn.
• Read dedup: a file is read once.
• RTK mode: optional, off by default.
• Measurement: twenty-five matched tasks, run both ways, to compare token usage.

Install

and integrate Install the CLI and runtime once, then run compression beside your application: npm install -g @caveman-ai/cli caveman setup --install CAVEMAN_MODE=compress caveman start TypeScript adapters are published as pinned alpha packages: npm install @caveman-ai/adapter @caveman-ai/core Caveman works with Claude Code, Codex, and Gemini CLI; the skill and the proxy are two separate installs. It supports Linux, macOS, and Windows.

Evaluation

and rollout Caveman provides a record, replay, shadow, canary, and active flow for evaluating an AI agent optimization before rollout. It also offers a team-wide analytics view and a tool catalogue covering every tool name, description, and schema.