Caveman makes coding agents talk like cavemen, cuts tokens 33%, 113k stars
techNmak · x · 2026-10-18
Caveman, an open-source project with over 113k GitHub stars, makes coding agents respond in caveman style: the same React re-render explanation drops from 63 to 20 tokens with identical diagnosis and fix. It also ships a local proxy that compresses what agents read — logs, test outputs, JSON/CSV/YAML, large tool responses, web content — with originals still retrievable, cutting session input tokens by 33.2% in the project's tests. JetBrains independently measured 8.5% fewer output tokens on real coding tasks with no statistically detectable quality loss.
More from coding & agent
- Substack MCP Server Ships for Creators: Drafts, Notes, Analytics and Subscriber Tools — modelcontextprotocol · 2026-10-18
- Packrift Launches Verified MCP for Packaging: Live Pricing, Stock and Ordering — modelcontextprotocol · 2026-10-18
- Dev builds voice-note-taking Grok Bot template that auto-organizes ramblings into to-dos — XFreeze · 2026-10-18
- Google's Deep Review tore a 35-page benchmark spec apart, finding 36 flaws — ChrisUniverse · 2026-10-18
- Dev open-sources an X network map tool with a Grok agent prompt to collect data — jacob_posel · 2026-10-18
- Serverless, explained with a pizza shop and ghost kitchen analogy — thisiskp_ · 2026-10-18