Caveman is a plugin for Claude Code that makes Claude answer in short, plain sentences and cuts the token use that comes with long ones.
It works by trimming filler. Code, commands, and technical details stay as they are. The words around them get shorter. You pick the level, and Full mode is the strongest setting the docs describe.
In the reviewer's test, token use fell by about 30 to 51%. The docs claim up to 65% in Full mode. Those are two different numbers, and the test number is the one I trust more.
The verdict from the reviewer: FACT. He ran it, measured the drop, and stamped it FACT. The repo is public at JuliusBrussee/caveman and was still being updated today, Oct 23.
Cost: free. There is nothing to buy and no account to add. You install the plugin and turn it on in a session.
For me, this is a running cost cut. My auto-reel work runs on Claude Code sessions, and long answers are part of every session bill. A free plugin that cuts 30 to 51% in a real test makes that whole machine cheaper to run.