Caveman
Products
Agent toolkitlive

Caveman Code

1.93× fewer tokens on 25 matched coding tasks

A terminal coding agent measured at 1.93× fewer tokens than Codex CLI across 25 matched gpt-5.5 tasks, with 14/25 versus 15/25 tasks passing. Four compression layers sit under an autonomous goal loop across 20+ providers.

Product demo / Caveman Code

scripted terminal run
caveman-code · scripted run
$caveman-code "add a /healthz route and a test"

token budget

messages2.1k / 30k
memory1k / 40k

layers active

caveman mode
tool budgets
read dedup
optional RTK

published 25-task bench · 1.93× fewer tokens · 14/25 vs 15/25 passed

Capability ledger

06
  1. 01Four compression layers: Caveman Mode, tool budgets, read-dedup, optional RTK
  2. 0225-task MicroBench: 524k vs. 1,010k fresh tokens; 14/25 vs. 15/25 passes
  3. 03Autonomous goal loop with autopilot, plus read-only plan mode
  4. 0420+ providers (Claude, ChatGPT, Copilot, Gemini, Vertex, OpenAI, Azure, Groq, DeepSeek…)
  5. 05Architect/editor model split, session branching, shadow-git checkpoints
  6. 06Persistent memory via cavemem, MCP servers, up to 7 worktree-isolated subagents

Product ledger

02
Language
TypeScript
License
MIT

Product index

05
01CavemanOutput compression for Claude Code & 30+ agents, plus the compression engine.Compression / live02CaveGemmaCaveman compression baked into Gemma's weights.Compression / live03CavememPersistent memory your agents recall over MCP.Agent toolkit / live04CavekitCompressed, spec-driven development.Agent toolkit / live05Caveman ProxyThe byte-safe LLM gateway.Cloud / in development