Caveman

Caveman starts from an uncomfortable fact: most of your AI bill is filler. Its stack attacks the waste from three angles — a Skill that strips filler from prompts while preserving intent, a Proxy that compresses logs, JSON, code diffs, tables and tool outputs before they hit the model provider (original bytes kept recoverable), and an Agent SDK for building on top. Caching, benchmark-gated model routing, token-usage visibility and dollar-savings proof are all part of the package, with MIT-licensed pieces and free local wrappers that use your own provider keys.\n\nThe "proves every dollar saved" framing is what separates it from vague cost-cutting advice: you see the before/after on your actual traffic, not a benchmark someone else ran.\n\nWho it's for: teams running agent-native products where token spend scales with usage, and individual devs watching their coding-agent subscriptions balloon. Model routing and the hosted platform are still in development, so today you're buying the compression and caching layer — which, for high-volume agents, is usually where most of the waste lives anyway.