Caveman Code is a terminal coding agent whose central idea is aggressive token economy: the model answers in terse technical fragments ("caveman" style) and every tool result is compacted before it enters context, so the same task on the same model consumes far fewer tokens. The README's own reproducible 25-task benchmark reports roughly 1.9x fewer tokens than Codex CLI at a comparable pass rate. It is a heavy fork of pi-code, tracked against upstream, with the compression layers, goal loop, and plan mode as its own additions.
Compression works in four always-on layers hitting two token sinks: Caveman Mode compresses model replies (lite/full/ultra levels), per-tool output budgets cap and clean shell/read/grep output, read deduplication returns stubs instead of bytes for unchanged re-reads, and an optional external RTK binary compacts bash output further. Aggressiveness is adjustable or can be switched off per session.
Around that core sits a full agent: an autonomous "Ralph-style" goal loop with per-iteration cost ledgers and shadow-git checkpoints, a read-only plan mode, up to seven worktree-isolated parallel subagents, an architect/editor model split, session branching and replay, YAML workflow recipes, a daemon mode for SSH-surviving sessions, and a TypeScript SDK. It supports 20+ providers via API keys or OAuth (including existing Claude, ChatGPT, Copilot, and Gemini subscriptions), speaks MCP, delegates persistent memory to the companion cavemem project, and deliberately keeps its settings, commands, skills, and agents formats as a superset of Claude Code's so existing configs work unchanged.
