Caveman
🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman
Last verified:
What is Caveman?
Caveman is a token-efficient stack for agent-native development built around a core compression primitive, a spec-driven workflow, a persistent memory layer, and a coding agent CLI. The site presents it as a three-part ecosystem, though it also describes four projects in the stack: Caveman, Cavekit, Cavemem, and Caveman Code.
The core Caveman primitive compresses prompts, system messages, long-form documents, and agent context into much shorter semantically equivalent forms. The site says it can cut token usage by about 75% on typical agent workloads, while remaining model-agnostic, deterministic, and composable. It is positioned as something you can plug in upstream of whatever model or agent you already use.
Cavekit extends that idea into a spec-driven workflow. It turns prose specs into structured plans and then into verifiable execution, with explicit tasks and acceptance criteria. The website emphasizes that it is opinionated, not clever, and that it uses Caveman internally so both the plan and the context stay lean.
Cavemem adds persistent local memory for coding agents. It stores compressed observations in local SQLite with FTS5 and vector search, exposes them through MCP, and lets different agents recall what previous sessions learned. Caveman Code is the flagship CLI that combines prompt compression, command compression, output compression, and context compression into one coding agent workflow.
This stack is aimed at builders and agent users who care about reducing token spend, latency, and context bloat. The site repeatedly frames it as useful for developers, especially those using Claude Code, Cursor, Codex, Gemini, or other model providers who want a leaner and more deliberate workflow.
Caveman pricing
Pricing model: Freemium
The website presents Caveman, Cavekit, Cavemem, and Caveman Code as open-source and MIT licensed, with no paid tiers or subscription pricing shown. Caveman Code is described as live on npm and installable with a global npm command, and the site also says it works with pnpm, yarn, bun, and Docker. No free tier versus paid plan breakdown is listed on the site; the pricing model shown is effectively free/open-source access.
Caveman pros
- Cuts token usage by roughly 75%
- Model-agnostic compression primitive
- Works on prompts, system messages, and documents
- Deterministic compression behavior
- Uses a controllable dictionary
- Composable with other tools and workflows
- No retraining required
- No proprietary runtime required
- Spec-driven planning in Cavekit
- Structured tasks with acceptance criteria
- Verification built into the workflow
- Persistent cross-agent memory in Cavemem
- Local-first storage with SQLite
- FTS5 and vector search support
- Exposed through MCP for interoperability
- Caveman Code compresses four layers independently
- Supports 20+ model providers
- Live npm installation for the flagship CLI
- Works with pnpm, yarn, bun, and Docker
- MIT licensed across the stack
Caveman cons
- Token savings are framed as typical, not guaranteed
- Compression mainly targets communication, not model thinking
- Best gains may be smaller in code-heavy sessions
- Cavekit is opinionated rather than flexible
- Cavemem requires local SQLite setup
- Cavemem depends on MCP integration
- Caveman Code is only the flagship layer, not the whole stack
- The stack still depends on your chosen model provider
Frequently asked questions about Caveman
What is Caveman?
Caveman is the compression primitive at the center of the stack. It takes prompts, system messages, long documents, or agent context and compresses them into a shorter version that keeps the meaning while reducing token usage.
How much token reduction does Caveman claim?
The site repeatedly says Caveman can reduce token usage by about 75% on typical agent workloads. It also shows an example where a prompt goes from 4,820 tokens to 1,204 tokens, which the site labels as a 75% reduction.
Does Caveman require retraining or a special runtime?
No. The site says Caveman does not require retraining and does not need a proprietary runtime. It is meant to sit upstream of the model or agent you already use.
What is Cavekit used for?
Cavekit is the workflow layer for spec-driven development. It turns written specs into structured plans, then into execution with verifiable tasks, so work starts from explicit goals instead of vague instructions.
How is Cavemem different from Caveman?
Caveman is about compression, while Cavemem is about persistent memory. Cavemem stores compressed observations locally in SQLite, uses FTS5 and vector search, and exposes the data through MCP so multiple agents can recall past context.
What is Caveman Code?
Caveman Code is the flagship coding agent CLI. It combines four compression layers: prompt input compression, reduced-token command handling, output compression through Caveman, and context or CLAUDE.md compression.
Which model providers does Caveman Code support?
The website says Caveman Code can point at any of 20+ model providers. It positions the CLI as model-flexible rather than locked to one vendor.
Is Caveman local or cloud-based?
The stack is presented as local-first. In particular, Cavemem is described as local SQLite storage with nothing leaving the box, and Caveman itself is described as a portable compression primitive that works wherever tokens are spent.
What installation options are shown for Caveman Code?
The site shows a global npm install command for Caveman Code and says it is also available through pnpm, yarn, bun, and Docker. It presents the flagship CLI as live and ready to install.
Who is Caveman for?
The site targets builders and agent-native developers who care about token cost, latency, and context bloat. It is especially framed for people using coding agents who want leaner prompts, leaner outputs, persistent memory, and a more structured spec-driven workflow.