JuliusBrussee/caveman

Health v2 · maintenance only

Flags: young no_license

Adoption not part of the score

101200 stars · 5871 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Caveman is an efficiency stack for AI coding agents that reduces token usage through a terse-writing skill, a payload-compression engine, and a local context memory layer. The MIT-licensed skill is a Markdown instruction file that makes agents drop filler while preserving code, commands, and error strings verbatim, with optional CLI tooling for profiling token spend.

Use cases

• reduce llm api token costs
• make claude code use fewer tokens
• compress prompts sent to model apis
• profile where agent tokens are spent
• cut verbose ai agent output
• lower context window usage in coding agents
• save money on anthropic api bills

When to choose

• you use Claude Code, Cursor, Codex, or similar coding agents and pay per token
• you want a zero-runtime, install-and-forget prompt skill that never alters code blocks or error strings
• you want local, recoverable payload compression with byte-exact restoration
• you want to profile which instructions and recurring context eat your token budget

When to avoid

• you need terse output to remain fully human-readable prose
• your workflow depends on the agent's explanatory narration or hedged phrasing
• you require a fully open-source stack including the memory and cloud layers (those are BSL 1.1 or commercial)
• you use an agent host not covered by the supported profiles and cannot tolerate generic fallback wiring

Facets

plugin · maturity active

prompt-engineering llm-inference cli developer-tools caching large-language-models developer-tools cli cross-platform python go token-optimization claude-code agent-skills prompt-compression context-compression cost-reduction open-core ai-agents command-line nodejs

Member repositories

For agents

markdown · JSON · MCP: product_card(name="JuliusBrussee/caveman")