Caveman promises 33.2% fewer tokens. We measured it and got 0.0% (JuliusBrussee/caveman)

We install caveman, the tool that hit #1 on GitHub Trending and #1 on Hacker News this year, and test the one number its 107,000 stargazers never checked: does the proxy actually cut tokens on a normal Claude Pro/Max login, not an API key. The free half is an MIT skill that makes your agent write like it just discovered fire, and it genuinely works. The other half is a local proxy claiming a 33.2% cut in input tokens. We run it against a real Claude Code session reading a synthetic four thousand line log with a single FATAL line buried in it plus a twelve hundred user JSON file, read caveman's own meter, get 0.0%, chase the fix the tool itself points at, and measure again. https://github.com/JuliusBrussee/caveman (107,280 stars) Chapters: 00:00 Welcome 00:04 Intro 00:17 The claim 00:58 The free half, and it works 01:56 The big rock 02:34 The measurement 03:14 The fix 04:15 Other findings 04:47 Verdict 05:52 Where you'd use it 06:16 Bottom line 06:28 Credits Honest verdict: the free skill is worth installing, it works as advertised. The proxy's 33.2% headline number didn't reproduce on a Pro/Max subscription login in this test, and the project's own docs already say savings on that path can be zero, so test it on your own workload before trusting the number. Recorded in an isolated container. Not sponsored.