Caveman
Dev Utilities Open source (Apache 2.0)
AuthenticationOpen source (Apache 2.0)
Rate limitSelf-hosted open source: no limits; reduces your model token spend.
Documentation: docs.caveman.so ยท Checked Oct 12, 2026
About
Caveman is a viral skill plus local proxy that claims to cut coding-agent token usage by up to 65% by making agents communicate tersely, like a caveman: short prompts, short replies, no ceremony. Whether it hits the full 65% on your workload or not, the technique is real and the proxy is Apache 2.0 open source. Worth trying if your agent bills are the bottleneck.
Details
Open source. You run it yourself.
Auth: Open source (Apache 2.0)
Rate limit: Self-hosted open source: no limits; reduces your model token spend.
Verified: 2026-10-12
Best for: developer automation
Setup
- Install the caveman skill and local proxy from the docs.
- Point your agent harness at the local proxy.
- Enable the terse-communication skill.
- Run a typical task and measure tokens before and after.
- Keep it on for the workloads where it wins.
Example
caveman proxy --harness claude-codeUse cases
- Reduce spend on high-volume agent runs
- Prototype cheaply before scaling up
- Pair with free inference gateways for near-zero-cost agents