
Caveman Code is a Mac and browser app running Claude, Codex, Cursor, and its own agent.
Where do rivals beat it, and what does each gap cost?
For input costs instead of reply costs, Headroom, RTK and SuperCompress serve as the alternatives, since Caveman's reply-style skill leaves input uncompressed, though its separate engine and proxy compress it.
JetBrains measured that the Caveman technique cut output token usage by roughly 8.5%, well under the 65% Caveman claimed.
Which risks could break the thesis, and what would warn of them?
The gap between advertised and measured savings is the risk most able to break Caveman's thesis: its paid case depends on measured savings holding on real agent work, and an independent test puts the saving near 8.5% of output tokens, about 10% of cost, at absolute best, with the total bill able to end higher. The advertised 65% saving belongs to chat-style question-and-answer rather than coding agents, and in the full run the skill arm's raw total was 11.6% more expensive, USD 40.60 against USD 36.39.
One dependency-audit task in the skill arm crossed the 200k long-context pricing tier and billed USD 8.29 against USD 0.33. Every saving Caveman reports is measured at the gateway and verified on live traffic before it counts, and none is claimed on a projection.
What does the product do for its buyer, and how is it delivered?
Caveman shrinks reading and writing by coding agents while keeping originals recoverable, and ties agent spend to work produced. The system places itself where tool results turn into model input, with the framework and provider retaining their separate functions; eligible logs, JSON, diffs and documents compress ahead of the model call, and in supported agent loops the model can request the exact original. Caveman Cloud tracks every coding agent a team runs and every agent it ships back to the work performed and its cost.
The agent and the SDK are available free, without an account. Caveman Code runs as a Mac app and in the browser, with Claude Agent, Codex, Cursor and Caveman operating concurrently in tabs and splits. Team pricing is $300 per month following a 7-day free trial.