AI Native Landscape

caveman

Tracked

A skill plus local proxy for coding agents that cuts token usage by making the agent answer in a compressed "caveman" voice and shrinking tool output before the model reads it, with measured savings and no loss of answer quality.

AuthorJuliusBrusseeOpen Sourced2026-04-04Last CommitUnknown

Overview

Caveman makes AI coding agents say less and read less while the code stays exact. One install (npx @caveman-ai/cli@latest) patches Claude Code, Codex, Gemini CLI, OpenCode, and 30+ other agents with six components: a compressed reply voice, pre-read compression of logs and tool output, automatic model and effort selection, reusable agent-written scripts, waste fixes, and compressed web browsing. External evaluations from Adobe Research, JetBrains, and Elastic back the token savings with no measurable quality loss.

Key Features

  • Say less: the agent answers in a structured short voice (63 → 20 tokens for the same fix, same code payload), with strict levels from /caveman down to /megacave
  • Read less: logs, JSON, CSV, and test output are shrunk before the model reads them — 33.2% fewer input tokens measured across 54 Claude Code runs
  • Auto: picks the model and reasoning effort per request, measured 54% cheaper than a maxed-out Opus configuration at the same task score
  • Blocks: the agent keeps scripts it writes and reruns them instead of re-reading output (397 → 76 tokens for one test run)
  • Waste fixes find where tokens go and fix it, including keeping the prompt cache warm while you pause
  • Browse returns web pages compressed (129.8x smaller page)

Use Cases

  • Cutting cost and latency of daily coding-agent sessions without changing how you prompt
  • Keeping subscription quotas and rate limits from interrupting long agent runs
  • Running the same agent workflow across Claude Code, Codex, Gemini CLI, OpenCode, and other CLIs with one setup
  • Teams wanting benchmarked token savings rather than blindly shortened answers

Technical Details

  • Installed via the @caveman-ai/cli npm package (Node 22.13+); the CLI detects installed agents and their logins and applies a prompt skill plus a local proxy layer that runs on your machine with your own keys
  • The voice is not broken grammar but a fixed reply structure based on ASD-STE100 controlled English: answer first, one idea per sentence, 20 words max, active voice; negations, numbers, and units are never dropped and code, commands, paths, and error text stay verbatim
  • User prompts are never rewritten — external research showed compressing the user's prompt backfires; compression targets only what the agent reads (tool output) and runs on (scripts)
  • The proxy compresses tool_result payloads (logs, JSON, CSV, test output) before they reach the model, translates web pages into compact form, and keeps the provider prompt cache warm across pauses
  • Each of the six components can be toggled individually (caveman status, caveman off browse); only the Auto model picker needs an account