Caveman Proxy
Your agent rereads context. Make it smaller locally
AI agents resend instructions, history, and tool results across a task. Caveman Proxy wraps agents you already use and sends that traffic through the local Caveman Engine, which recognizes logs, JSON, tables, code, diffs, and other common context. Original bytes are stored locally before lossy replacement so agent can retrieve exact detail later. Local compression needs no Caveman account. Released — get it from the caveman repo.
Product demo / Caveman Proxy
request-path model · illustrativeThe proxy is a drop-in base URL. It reads every request, adds provider-native cache hints to the upstream request only, and never alters the bytes your model sees. In record mode it is pure pass-through.
What it changes is what you can see: provider-reported usage and public-catalog list-price subtotals where coverage is complete. Unknown models remain visibly unpriced rather than displaying $0 as a known price.
telemetry
- model
- claude · gpt · gemini
- tokens in / out
- 1,284 / 312
- cache hint
- upstream only
- price
- catalog · or unpriced
- mode
- record · pass-through
demo fixture · no savings claimed before measured active traffic
Capability ledger
06- 01One-command launch for Claude Code, Codex, Gemini, Aider and other registered agents
- 02Content-aware: matching compressors for logs, JSON, tables, code, diffs, terminal output and more
- 03Recoverable: original bytes are stored locally before lossy replacement
- 04Fail-closed: parse failure, missing recovery or non-smaller output passes through unchanged
- 05Local token reductions are estimates labeled inferred, never verified savings
- 06Free local compression without account; powered by Caveman Engine
Product ledger
03- Language
- Go
- License
- BSL-1.1
- Pricing
- Local compression is free without an account. Cloud analytics, teams, governance and verified savings are commercial.