caveman: cut agent token use by 65%

TL;DR: caveman presents itself as a skill plus a proxy for coding agents, claiming a 65% cut in token usage by making the agent answer in clipped caveman speech. It is Apache-2.0 licensed, has 109,113 stars and 6,316 forks, and was pushed today.

What the repo claims

The description is a joke with a number attached: “why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.”

Take the bit out and the architecture is fairly specific. Two components, both named in the description. A skill, which is instruction text the agent loads, and a proxy, which sits in the request path. The savings are claimed on the output side, not through better retrieval or smaller context.

That distinction matters when you evaluate the claim. Trimming what the model writes back is the cheapest kind of token reduction and the least likely to hurt task quality, but the saving is bounded by how much of your spend is output. If your bill is dominated by large file reads and long tool traces, a shorter reply style will not move it much.

The numbers

Everything here comes from the GitHub API on the morning of writing, not from the README.

Field	Value
Language	Go
License	Apache-2.0
Stars	109,113
Forks	6,316
Open issues	134
Created	2026-04-04
Last push	2026-10-03
Topics	ai, anthropic, caveman, claude, claude-code

The topics are the clearest signal about who this is for. Anthropic and claude-code sit next to a generic ai tag, which puts the target squarely on Claude Code and compatible harnesses rather than on model-agnostic tooling.

A fork count worth noticing

Six thousand three hundred sixteen forks against 109,113 stars is a ratio of roughly one fork for every seventeen stars. For a library you import, that is high. For a repository whose main artifact is a set of files you copy into your own configuration, it is exactly what you would expect, since a fork is how you install it without depending on the upstream.

That reading fits the project type. A skill is a directory of text. You do not import it, you copy it into a path your agent reads at startup. The proxy is the piece that actually needs the Go binary.

The caveat is that the same affordance means forks can drift from upstream silently. If you install by forking, you own the maintenance.

How to try it

Start with the quickstart, which is the project homepage and the only setup documentation the repository metadata points to: docs.caveman.so/docs/quickstart.

Then check the numbers yourself instead of trusting a badge, since star counts move:

gh api repos/JuliusBrussee/caveman \
  --jq '{stars: .stargazers_count, forks: .forks_count,
         license: .license.spdx_id, pushed: .pushed_at}'

If you want to measure the 65% figure on your own workload, the honest method is to run a fixed set of identical tasks with and without the proxy and compare total input plus output tokens for each run. A percentage that arrives without a methodology in the description is a marketing number until your own numbers reproduce it.

When to skip it

Skip it if your agents write code that humans review in a diff. Compressing the assistant’s narration to save tokens is a reasonable trade when the output is discarded and a bad one when the output is a patch a colleague has to read at 9pm. The skill changes how the agent talks, and in some workflows that text lands in the pull request description or the commit message.

Skip it if you cannot run a local proxy in the request path. The design puts a process between your harness and the model API, which is fine on a laptop and not always fine inside a locked-down network or a CI sandbox.

Source and Accuracy Notes

• Project page: https://github.com/JuliusBrussee/caveman
• Source repository: JuliusBrussee/caveman (verified via GitHub API html_url)
• License: Apache-2.0 (verified via GitHub API license.spdx_id)
• Stars at time of writing: 109,113
• The 65% token reduction is a claim made by the project in its own repository description. It was not independently reproduced for this post.