JuliusBrussee/caveman

The problem

Agent sessions burn tokens in places a "be concise" instruction never reaches: verbose prose answers, entire log dumps and JSON payloads fed back into context, and repeated re-reads of the same command output. If you pay per token or hit context limits mid-task, the visible cost is the answer text, but the bulk sits in what the agent ingests. Caveman is a skill plus local proxy that installs into an existing coding agent and shrinks both sides.

What's different

Most token savers are either a prompt instruction ("/caveman" alone lives here as a skill) or a generic LLM gateway that sits between you and the API and cannot see tool calls. Caveman installs into the agent runtime across 30+ agents, so it can compress tool output and logs before the model reads them — the README's claim of 33.2% fewer input tokens per session comes from that layer, not the voice. Three capabilities stand out: a structured short-answer style built on ASD-STE100 controlled English that leaves code, commands and error strings byte-exact; a Blocks code mode that lets the agent keep and re-run scripts instead of re-deriving output (397 → 76 tokens for one test run, per the README); and an Auto router that picks model and effort per request. Notably it deliberately never rewrites your own prompt, citing outside research that this backfires — a more disciplined stance than most prompt-trimming tools.

Who it's for

Developers running long agent sessions on metered API keys, or anyone repeatedly blowing context windows on test output and logs. Also useful for teams paying per seat where shorter answers still reduce context pressure. Look elsewhere if you use GitHub Copilot-style per-request billing, where the README itself admits shorter answers save nothing, and skip it if you need auditable, deterministic prompt handling — a proxy that rewrites inputs sits in your trust path.

Cost of adoption

Needs Node.js 22.13+; install is npx @caveman-ai/cli@latest, with shell and PowerShell one-liners for wider agent coverage. Apache-2.0. It runs locally with your keys and your Claude Pro/Max login. Risks: the voice rock costs about 1,160 tokens of rules input per session, so gains depend on session length; only Auto needs a (free) account; the README's own numbers are partly interim and some rocks are explicitly unmeasured. External validation exists (Adobe Research paper, JetBrains A/B, Elastic writeup), which is unusual for this category and worth weighing.

Verdict

This is a 203-day-old project with roughly 114.8k stars, but the trend board shows it gaining only a couple hundred stars a day on two recent days — momentum has cooled from its April HN spike (904 points) into maintenance mode. The honest-numbers file and external citations suggest the maintainers care about claims being checkable, which is rare for a meme-branded token saver. Treat the headline 65% as a bundle of separately measured pieces, each with its own caveats. Worth trying the free skill first to see if shorter answers fit your workflow, then deciding whether the proxy layer earns a place in your toolchain.