caveman

# caveman

A collection of tools that minimizes token consumption for AI agents by compressing logs, managing agent scripts, and optimizing responses. It is designed for developers who want to reduce API costs while maintaining agent performance.

Made by juliusbrussee on GitHub. These notes are Minda's, written from its README; the project's own files and license are in its repository.

## About

What it is: Caveman is a set of tools designed to reduce token consumption by AI agents. It compresses logs, manages agent scripts, and optimizes agent output to lower costs while keeping code and information exact.

Who it suits: It is for developers and engineers who use AI coding agents and want to manage their token usage.

Where to use it: You can use it within terminal-based coding agents, CLI tools, and various development frameworks.

- Claude Code
- Gemini CLI
- Codex
- OpenCode
- Vercel AI SDK
- LangChain

How to set it up: You need a machine running Node.js 22.13 or newer on Mac, Linux, or Windows.

1. Install the CLI using the command provided.
2. Run the tool to automatically find agents and logins.
3. Use the status command to see which features are active.

```bash
npx @caveman-ai/cli@latest
```

How to use it: You can interact with the tools by running commands in your terminal or by invoking slash commands directly through your agent's chat interface.

```bash
caveman status
```

| Tool | What it does | |---|---| | /caveman | Sets the agent to respond with short, direct phrases | | /caveman-compress | Shrinks memory files while keeping the content intact | | caveman-blocks | Manages and reuses common scripts written by agents | | caveman login | Connects to a free account to enable model routing |

Example: Instead of a long explanation, the agent provides a direct fix. The answer remains accurate but consumes fewer tokens.

For newcomers and experts: Newcomers can think of this as a filter that asks the AI to be brief, which saves time and money. Experts can use it as a token-efficiency layer that sits between the agent and the model provider. It uses response-voice control, log-file compression, and script-block reuse to optimize token throughput. It trades off conversational style for structured output that preserves functional integrity.

Good to know:

- The tool operates locally on your machine for most tasks.
- It can read your local files, logs, and agent scripts.
- It is released under the Apache-2.0 license.
- Features like Auto model routing require a free account and may send metadata to the cloud.

```chart
Title: Average token savings
Input tokens saved: 33.2 percent
Response tokens reduced: 8.5 percent
```

## On GitHub, 5 November 2026

- Stars: 117,527
- Forks: 6,820
- Watching: 253
- Open issues and pull requests: 157
- Language: Go
- License: Apache-2.0
- Topics: ai, anthropic, caveman, claude, claude-code, llm, meme, prompt-engineering, skill, tokens
- Trending: #1 repository of the day on GitHub trending (5 November 2026)

## Links

- Code: https://github.com/juliusbrussee/caveman
- Website: https://docs.caveman.so/caveman
- Releases: https://github.com/juliusbrussee/caveman/releases
- Issues: https://github.com/juliusbrussee/caveman/issues
- Star history: https://www.star-history.com/#juliusbrussee/caveman&Date

## Intro videos

- No way this actually works: https://www.youtube.com/watch?v=L29q2LRiMRc

## Get the code

```
git clone https://github.com/juliusbrussee/caveman.git
```

Then follow the setup in its README. Nothing in this kit runs by itself.

README.md  3 KB