Caveman

What is Caveman?

Caveman is a token-efficiency stack for AI coding agents, sold as layers you can adopt separately. The free layer is an MIT-licensed skill that makes agent output terse plus a local proxy (`caveman wrap claude`) that compresses provider traffic on your machine — the Caveman Engine recognizes logs, JSON, code, diffs and tables, stores the original bytes, and sends a smaller eligible context to the model, with no account required. Paid layers add a hosted gateway, a spend workspace that files every dollar under a cause, and a TypeScript agent SDK. Its pinned benchmarks report 33.2% fewer provider-reported input tokens through the proxy and around 65% fewer output tokens from the skill.

Key features

• MIT-licensed skill that rewrites agent output in terse 'caveman' style, installable into Claude Code, Codex, Cursor and 30+ other agents
• Local Caveman Proxy that wraps an agent's provider traffic (`caveman wrap claude`) and compresses eligible context before it is sent — no Caveman account needed
• Recoverable compression: the Engine recognizes logs, JSON, code, diffs and tables and stores the original bytes before lossy replacement
• Spend workspace showing who spent what, through which key, on which model, and why — broken out by cause (uncached prefix, oversized tools, premium on easy work, retries)
• Caveman Agent SDK (`@caveman-ai/agent`, TypeScript) that declares tools, context and evals, and only keeps a cheaper context plan after the declared checks pass; locked adapters for Vercel AI SDK, Eve and Mastra
• Hosted gateway and savings dashboard as an optional cloud tier, with the local wrap remaining usable on its own
• Enterprise on-prem option where the Caveman control plane receives no prompt bytes and zero data retention is enforced at write time
• Split licensing: the skill and the browser extension are MIT; the Caveman Engine and proxy runtime are BSL-1.1 (source-available, not OSI open source)

Who it's for

• Cutting the output half of a coding-agent bill by installing the free MIT skill into an agent you already run
• Compressing large logs, tool output, diffs and files before they enter a provider call, via the local proxy
• Attributing agent spend across a team — per person, key and model — to find which behaviors (retries, oversized tools, uncached prefixes) are burning the budget
• Building a production agent in TypeScript with a per-agent token bill and eval-gated context plans through the Agent SDK
• Running compression on-prem where the control plane must not receive prompt bytes

When not to use it

Skip it when you need savings you can reconcile against a provider invoice or an unmodified prompt payload: the free local run reports token reductions estimated locally rather than measured against a bill, the compression is lossy-with-recovery on the way out, and only the skill and browser extension are MIT — the Engine and proxy runtime are BSL-1.1, so it is not a drop-in for policies that require OSI-approved open source.

FAQ