Caveman is an open-source toolkit for cutting the token cost of AI coding agents. It began in April 2026 as a skill that makes agents such as Claude Code answer in terse, "caveman" style prose, and has grown into a CLI, an SDK and a hosted cloud. The project is maintained by Julius Brussee and released under the Apache 2.0 license; the core repository is written mainly in Go.
Version 4.0 (October 2026) organizes the toolkit into six parts. Say less keeps agent answers short while code stays exact. Read less shrinks logs, JSON, CSV and test output before the model reads them. Auto picks the model for each request and requires a free Caveman account. Blocks is a code mode in which the agent keeps the scripts it writes and runs them again; it is still pre-release. Waste fixes show where tokens go and keep the prompt cache warm during pauses, and Browse returns web pages compressed. Caveman says it works with 30+ coding agents, including Claude Code, Codex, Gemini CLI and OpenCode.
Installation is a single command: npx @caveman-ai/cli@latest (Node.js 22.13 or later, macOS, Linux and Windows). The skill alone can be added with npx skills add JuliusBrussee/caveman -g, and there are a Claude Code plugin, a Gemini CLI extension and install scripts for shell and PowerShell.
Savings figures are mostly vendor-reported: Caveman cites about 65% fewer output tokens for the skill and 33.2% fewer input tokens for Read less, and Auto's cost results are labelled interim. Independent write-ups by JetBrains and Elastic and an Adobe arXiv paper also measured shorter responses, so treat the exact percentages as workload-dependent.
Caveman Cloud adds LLM observability, agent runs and routing for teams building their own agents. It is free for one person, and the Team plan costs $300 a month with $300 of usage included and no per-seat fees. Enterprise runs in your own VPC. Caveman states a 0% markup on tokens; model calls use your own provider API keys.
Data handling differs by component. The skill sends nothing. The CLI sends pseudonymous usage telemetry by default (commands from a fixed list, counts, OS, install ID; no prompts, file paths or credentials), which you can disable with caveman telemetry off or DO_NOT_TRACK=1. With routing on, recent messages are sent to Caveman Cloud to pick a model. On the Cloud Free plan, content that reaches Caveman Cloud may be used to improve Caveman's own models; you agree to this at sign-up.
