Skip to main content
AdOpenFree logoPromote your productReach more potential users and drive product growth and revenue.Advertise
Favicon of Caveman

Caveman

Free Listing

Caveman is an open-source response skill and local CLI proxy for coding agents. It shortens answers and compresses inputs like logs, JSON, CSV, and test output while keeping code, commands, paths, and error messages exact.

Visit JuliusBrussee/caveman

Overview

Caveman is an open-source project for coding agents with two main parts: a response skill that makes agent answers shorter and a local proxy that compresses what the agent reads. The skill follows a controlled, answer-first style while preserving negations, numbers, units, code, commands, paths, and error messages. The proxy runs on the user's machine with the user's keys and keeps every original on disk so the agent can retrieve it.

Key Features

  • Response skill shortens agent answers with an answer-first style and preserves negations, numbers, units, code, commands, paths, and error messages.
  • Local CLI proxy compresses logs, JSON, CSV, YAML, and test output before the model reads them; the benchmark reports 33.2% fewer input tokens across whole sessions.
  • Works with Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, and 30+ more agents.
  • Provides /caveman, /ultracave, and /megacave modes plus commands for commit messages, code review, memory-file compression, stats, and help.
  • Includes optional Auto model and effort routing, Blocks code mode, browse compression, and middleware for TypeScript or Python applications.
  • Runs locally, uses the user's keys and Claude Pro/Max login, and keeps original files on disk for retrieval.

Use Cases

  • Reduce input and output token usage in coding-agent sessions without changing code payloads.
  • Compress noisy command output, logs, test results, JSON, CSV, YAML, or web pages before an agent reads them.
  • Shorten responses in supported assistants such as Claude Code, Codex, Gemini CLI, Cursor, and Copilot using caveman response modes.
  • Shrink memory files such as CLAUDE.md while keeping headings, code blocks, and paths intact.

Getting Started

  • Install the full tool with npx @caveman-ai/cli@latest, then start an agent with caveman claude or another supported command. For the response skill alone, run npx skills add JuliusBrussee/caveman -g and type /caveman if it does not start automatically. The CLI requires Node.js

Deployment & Requirements

  • Node.js 22.13 or newer.
  • Mac, Linux, or Windows.
  • Runs locally with your keys and Claude Pro/Max login; no hosted account is needed for local use.
  • Auto requires a free account.

Before You Adopt

  • License: Apache-2.0. Review its terms before using, modifying, or distributing the project.
  • The CLI requires Node.js 22.13 or newer.
  • HTML currently has no compressor, so that file type may not benefit and can incur overhead.
  • Blocks is pre-release, Python-only, and does not yet include a sandbox.
  • Auto routing requires an account, though the account is free.
  • Per-request billing can make shorter answers offer no cost saving.

Comments

Sign in to leave a comment.

More like Caveman

Favicon of Taste Skill

Taste Skill

Free ListingStars: 100.5K

Agent skills that give coding agents taste

Coding Agents

Taste Skill is an open-source Agent Skill that gives coding agents art direction, reference images, and a real-browser check so they build distinctive websites and app screens instead of generic AI pages.

Favicon of ZCode

ZCode

Free ListingStars: 8.1K

AI coding harness for complex goals with multiple agents

Coding Agents

ZCode is Z.ai's AI coding workbench and coding agent harness for complex goals, offering desktop, web, and terminal experiences, multiple agents, GLM-5.3 integration, and remote control from WeChat, Feishu, or Telegram.

Favicon of Codewhale

Codewhale

Free ListingStars: 41.1K

Open-source coding agent that works with any model.

Coding Agents

Codewhale is an open-source coding agent and terminal client that reads projects, edits files, runs commands, checks its work, and connects to hosted or local models.