JuliusBrussee/caveman
Caveman compresses agent inputs and outputs, with local tooling for noisy commands, browsing, memory, and skills. Its own caveat matters: terse workloads can cost more once the skill prompt is included.
Caveman combines a terse-response skill with a local proxy that compresses agent inputs. A pinned Claude Code test reports **33.2% fewer input tokens** across 54 runs while passing 18 exact-answer checks; the original skill claims **65% fewer output tokens** across 10 tasks.
Measure it on your own sessions before adopting it. The CLI can inspect local history, rank token sinks, trial changes, compress command output, and revert edits that fail to lower tokens per turn.
Caveman combines a terse-response skill with a local proxy that compresses agent inputs. A pinned Claude Code test reports **33.2% fewer input tokens** across 54 runs while passing 18 exact-answer checks; the original skill claims **65% fewer output tokens** across 10 tasks. Measure it on your own sessions before adopting it. The CLI can inspect local history, rank token sinks, trial changes, compress command output, and revert edits that fail to lower tokens per turn. The output skill adds **about 1–1.5k input tokens per turn**, so already-terse work can regress. The engine-linked runtime is BSL-1.1 rather than OSI open source, and anonymous command and token-count telemetry is enabled by default but can be disabled.
Caveman adds direct, reversible token optimization to the candidates’ broader context-management approaches: it measures session-level sinks and can test both input compression and terse output rules. Its mixed economics narrow the lesson—compression is workload-dependent, and the skill’s recurring prompt cost can outweigh savings—so adoption should follow local exactness and tokens-per-turn trials rather than headline reductions.