Caveman is a coding-agent tool that trims per-turn reads and writes, picking the cheapest model that completes each task.
Research Notes
Caveman's compression approach is relevant to repeated context waste in coding-agent sessions, but sustained investment appeal depends on savings persisting at the whole-task level. Coding agents spend tokens on verbose logs, entire files read to handle one function, and repeated context across turns. Caveman shrinks logs, test output, JSON, and diffs before the model reads them, with originals available on request.
Output-token savings were 8.5% over 82 paired tasks rather than the advertised 65%. Headroom and Caveman both use a local proxy to compress agent context before it reaches the model.
Caveman serves developers who run coding agents or build apps. It trims what each agent turn reads and writes, routes each task to a model, and runs on the user's machine with the user's keys.
Both Headroom and Caveman route agent context through a local proxy that compresses it before the context reaches the model.
The extent of savings for agentic coding work depends on the cost of each task and how much output can be shortened. In JetBrains' forced-Caveman benchmark of 82 paired agentic tasks, Caveman achieved output-token savings of only 8.5% against an advertised 65%, and in the full run its skill arm cost 11.6% more than baseline, with one dependency-audit task priced at USD 8.29 versus USD 0.33.
Across the 82 paired tasks in JetBrains' full run, the arms were statistically indistinguishable in quality.