Caveman

Key advantages

• MIT license for the core (the skill itself, the Agent SDK, the CLI, the client SDKs), so it's freely modifiable and embeddable
• Two install tiers, from the skill alone up to the full proxy with Claude Code hooks and a statusline
• Self-reported 33.2% token reduction on a pinned Claude Code benchmark published in the repo (docs/WRAP-BENCHMARK.md), passing all 18 exact-answer checks in that test

Who it's for, and why

Developers who use Claude Code daily and want to cut token spend on long sessions without sacrificing answer accuracy.

• Cutting the volume of text an agent generates during long debugging or code review sessions
• Compressing what the agent reads before every API call via the Proxy component (beyond Claude Code: Codex, Gemini, Aider, OpenCode...)
• Adopting a terse, filler-free output style as a working principle, with or without installing the package

Getting started

1. 1Install the skill alone (MIT core): `npx skills add JuliusBrussee/caveman`.
2. 2For read-side compression before every provider call (the more powerful Proxy product): `npm install -g @caveman-ai/cli && caveman setup --install`, then `caveman claude` (or `codex`, `gemini`, `aider`, `opencode`, `hermes`, `openclaw`).
3. 3A full installer with Claude Code hooks and a statusline also exists, via a `curl | bash` script pinned to a version tag (v2.3.1 at the time of writing), for anyone who wants the deepest integration.
4. 4Full documentation at docs.caveman.so/docs/quickstart.

Things to watch out for

• Split license: the skill (the core, the Agent SDK, the CLI, the client SDKs, the contracts, the provider catalog, the extension shell) is MIT, but the Proxy/Engine/Cache Engine/rewriter/Browse/MCP server component is BSL-1.1, source-available but not OSI-approved open source, auto-converting to Apache-2.0 by June 21, 2030 or four years after each version ships, whichever comes first. The more powerful Proxy product is therefore not fully open source.
• The 33.2% token-reduction figure is self-reported by the project on its own Claude Code benchmark (docs/WRAP-BENCHMARK.md), not independently audited: treat it as the maintainer's claim, not a verified third-party measurement.
• This site applies the filler-free output principle in its own Claude Code sessions, but hasn't installed this specific package as a separate plugin: this entry documents the community tool that popularized the technique, not a dependency actually installed on this project.