Caveman

AI coding agents running in autonomous loops accumulate token costs fast: every re-sent context window, every redundant plan, every forgotten prior decision compounds. Caveman addresses this with five stacked layers — a compression proxy, persistent agent memory, a terminal coding agent, a telemetry-driven planning tool, and an eval-gated rollout system. Installation is local-first: a curl script for the Claude Code skill, npm packages for Caveman Code and Cavemem, and a browser extension for ChatGPT, Claude, and Gemini contexts. The vendor states the system is byte-safe — the model-visible bytes are never altered, only the billed token count changes.

The differentiating feature is Cavemem, a persistent recall layer served over MCP. Rather than re-injecting full conversation history into each request, agents query a local SQLite database with FTS5 and a vector index to pull only what is relevant. This is the architectural answer to one of the most expensive failure modes in multi-session agent work: agents that have no memory of what they already resolved.

Caveman Code, the terminal agent, adds a planning phase before any code ships, running across 20+ providers in a single autonomous loop. Cave Architect sits above it, turning measured token spend into a ranked list of optimization moves, each tagged with the application change it requires — splitting the decision of what to fix by how much it costs to fix it. Eval-gated rollout then gates any claimed savings behind replay, shadow, and canary stages; the vendor states nothing counts until the eval passes and auto-rollback is armed if quality slips.

The cloud Caveman Proxy — which would compress any LLM traffic via a single base URL swap — is described on the page as ‘coming soon’ with a waitlist. The open-source components (MIT license) install today; the gateway does not. Teams needing organization-wide token metering across multiple services cannot use that layer yet.