caveman-code

Why we included this project

Token spend is usually the first thing that eats a coding agent budget, and this project attacks the problem from an unusual angle: rather than swapping in a cheaper model, it wraps the model you already have and gets the assistant to reply in terse technical fragments, capping chatty tool output like grep results and git diffs before they re-enter the context window. Savings compound because the agent re-reads its own history on many turns, and the author publishes a reproducible benchmark putting it at roughly half the tokens of Codex on the same tasks. Be aware the repo is frozen as of August 2026, so it is a stable, finished tool rather than an evolving one, and follow-up work lives in the successor project, whose wrap mode makes the agent you already use cheaper instead of replacing it. A practical bonus for teams with existing subscriptions: you can sign in with Claude Pro or Copilot rather than setting up a separate API key.