VarwoCaveman GitHub: Agent Skill and Token Claims
Locate Caveman, choose a setup and understand what token-saving claims measure.
JuliusBrussee/caveman is a collection of tools and agent skills for more compact AI workflows. The response skill encourages shorter prose; the project also describes other components such as local compression and routing. Those are distinct mechanisms with different setup and permissions.
Choosing a starting point
Read the repository’s current installation documentation. Decide whether you want only a response skill or the broader runtime before granting installation or configuration changes. Review the files and the licensing notes: the repository states that releases from 3.0.0 use Apache-2.0, while older versions and third-party components can have different terms. Caveman Cloud is a separate commercial service.
Do shorter answers always save money?
No fixed saving is guaranteed. Output length is only one part of a bill; input, caching, model pricing and the extra instructions matter too. The maintainers publish measurements and their limitations, including comparisons against concise baseline instructions. Compare the same representative tasks, model, settings, correctness and total usage in your own environment.
The project’s own technical notes say the response skill does not itself compress request context. Documentation checked November 4, 2026. Varwo has not installed it or measured savings.
Skill or runtime?
| Choice | What to inspect |
|---|---|
| Response skill | Instructions that change the agent's prose; compare with a concise-prompt baseline |
| Runtime and other components | Configuration, network paths, supported agent and applicable licence for your version |
Before accepting an installer, inspect what it will add or change. Saving output tokens does not prove that context was compressed, and a shorter answer can omit useful information.