Caveman
About
Open-source skill that makes AI coding agents talk like a caveman, cutting ~65% of output tokens. Works with Claude Code, Codex, Gemini, Cursor, and 30+ agents.
Our Verdict
Caveman is a genuinely useful token-saving skill for AI coding agents — it compresses output without sacrificing accuracy, backed by independent benchmarks from JetBrains. The 65% reduction on chat-style replies is real, though the 8.5% on agentic coding runs is more modest. The /caveman-compress feature for memory files is a clever addition that saves input tokens permanently. Best for developers who use Claude Code, Codex or similar agents heavily and want to cut costs without changing providers.
Pricing
Use Cases
- Reducing API costs when using Claude Code, Codex or other paid AI coding agents
- Speeding up code review by getting concise, one-line PR comments
- Compressing CLAUDE.md and other memory files to save input tokens every session
Pros
- Reduces output tokens by ~65% on prose, 8.5% on agentic coding runs (JetBrains verified)
- Works with 30+ agents: Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot and more
- Six compression levels (lite/full/ultra/wenyan) switchable mid-session
- Includes /caveman-compress for memory files (~46% input token savings forever)
- Free and open source with MIT license
- No quality degradation — independently verified by JetBrains on 86 auto-graded tasks
Cons
- Only compresses output tokens — input tokens unchanged, and the skill adds ~1-1.5k input tokens per turn
- Whole-session savings vary significantly by workload; on already-terse code runs savings can be minimal
- Install command (curl | bash) is a security concern for some users
- Project is young (April 2026) with 470 open issues
- Need Node ≥18 installed on the machine
Installation
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bashFrequently Asked Questions
What is Caveman?
Caveman is an open-source skill for AI coding agents that makes Claude Code, Codex, Gemini and others "talk like a caveman" — stripping filler and pleasantries from replies, keeping only the technical substance. Benchmarked at ~65% average output token reduction on prose, and 8.5% on long-horizon agentic coding runs (JetBrains, 86 tasks) with no statistical quality difference.
How do I install Caveman?
One command: curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash. It auto-detects every supported agent on your machine (Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, etc.) and installs for all. Activate with /caveman or "talk like caveman". Per-agent install also available: claude plugin marketplace add JuliusBrussee/caveman.
Is Caveman free? Does it really save money?
Completely free and open source (MIT). Savings depend on workload: chat-style replies (explanations, reviews, debugging) save ~65% output tokens; autonomous coding runs (JetBrains, 86 tasks) save ~8.5%. Caveman only compresses output tokens — input tokens unchanged, and the skill adds ~1-1.5k input tokens per turn. Whole-session savings vary. The /caveman-compress command also shrinks memory files like CLAUDE.md by ~46%, saving input tokens forever.