Reduce coding agent token costs by shortening responses and compressing inputs
The tool makes agents answer concisely instead of providing verbose explanations and compresses logs or data they read. This allows the same coding tasks to be completed at a much lower cost. It is a collection of proxies and skills designed to reduce the tokens consumed by AI coding tools.
Six features in version 4.0
A recent update to version 4.0 bundles six features into one. These are 'Say less', which shortens agent responses; 'Read less', which reduces input data; 'Auto', which automatically selects the model and effort; 'Blocks', which reuses scripts; 'Waste fixes', which identifies wasted tokens; and 'Browse', which compresses web pages. Each feature can be toggled independently, and all are enabled by default after installation. In particular, the 'Auto' feature saves costs by choosing a large model for difficult tasks and a small model for easy ones.
Measured savings
According to research from Adobe Research, this approach reduces costs by 1.4 to 2.4 times per model, up to a maximum of 3 times. In an A/B test by JetBrains, output tokens decreased by 8.5% without a drop in quality. Elasticsearch Labs reported a 63.6% reduction in response tokens with no loss of information. Analyzing entire Claude Code sessions showed a 33.2% decrease in input tokens. These figures are based on 85 real coding tasks and 54 runs.
Installation and requirements
It can be installed with the single command npx @caveman-ai/cli@latest in an environment with Node.js 22.13 or higher. Mac, Linux, and Windows are all supported, and the installation process automatically finds and configures existing agents and login information. It is immediately usable in more than 30 agents, including Claude Code, Codex, and Gemini CLI. If you are already using it, you must specify the @latest tag to update to the newest version.
Things to check before use
The tool is released under the Apache-2.0 license and can be used freely. However, using the 'Auto' feature requires a free account, and request content is sent to Caveman Cloud at that time. If privacy is a priority, you can turn off routing with caveman off routing or disable telemetry with caveman telemetry off. HTML file compression is not yet supported, and some file types may actually increase token usage, so you should test it in your own work environment.
