juliusbrussee/caveman is a Go tool that shortens the input material and responses of coding agents to reduce token usage. It provides skills and a proxy, and the latest release is v4.0.1. The repository description says it cuts tokens by 65% by changing how things are phrased. However, the provided information does not list anything newly added or changed in v4.0.1.
The tool bundles features that condense agent responses and compress logs, JSON, and exam output. It lets you pick the model and reasoning effort per request, and it also documents reusing scripts you have written and compressing web material. It can be installed from the command line and used with multiple coding agents, and the README explains that the processing runs on your own computer. The Go implementation and the Apache-2.0 license are also published. In short, it is an approach to reducing the inputs and responses exchanged between agents and models.
The repository explains that it separates its own measurements from external research and use cases, and it also marks features it has not yet measured. But from the provided information alone, neither the measurement conditions behind the 65% figure nor the concrete changes in v4.0.1 can be confirmed. So rather than applying that figure to every feature and environment, it is more reasonable to read it as a claim made by the repository, limited to that, until the scope of measurement is verified.