</>MCP Agents Market
MCP Server

caveman

by JuliusBrussee98.5kGoUpdated 2026-08-16

🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman

Claude CodeCodexGemini CLICursorWindsurfClineAider

Caveman is an MCP server and CLI tool that reduces AI agent token consumption by up to 65% through compression techniques and concise output formatting. It includes two components: a skill that makes agents respond in terse "caveman-speak" while preserving code accuracy, and a local proxy that compresses input payloads (tool schemas, files, logs) before sending them to language model providers. The tool works with Claude Code, Codex, Gemini CLI, Cursor, Windsurf, and 30+ other AI coding agents, requiring no code changes to integrate.

Key Features

Dual compression system: output skill reduces agent verbosity by 65% on average, proxy cuts input tokens by 33.2% in benchmarks
Content-aware compression routes JSON, logs, code, diffs, and search results through specialized compressors with byte-exact recovery
Pixel mode renders dense text blocks as PNG images for vision models, reducing tokens by up to 79% on minified JSON and long logs
Automatic token usage analysis with 'caveman learn' command that scans agent history and identifies top token sinks with actionable fixes
Browser automation compression shrinks web page accessibility trees by 129.8× (15,704 to 121 tokens) using focused queries
CCR (Content-Addressable Recovery) store keeps original bytes locally so agents can retrieve uncompressed data when needed
Native wrapper profiles for seven agents (Claude Code, Codex, Gemini, Aider, opencode, Hermes, OpenClaw) plus universal baseURL integration
MCP tools exposed to any host: caveman_compress, caveman_retrieve, caveman_stats, caveman_toon_encode, caveman_toon_decode

Use Cases

  • 01Reducing token costs for developers running frequent AI coding sessions with agents like Claude Code or Cursor
  • 02Compressing noisy command output, test logs, and build artifacts that bloat agent context windows
  • 03Converting installed skill documentation to compressed PNG format to reduce per-invocation overhead
  • 04Analyzing historical agent sessions to identify and fix top token consumption sources across projects
  • 05Enabling longer context retention by shrinking tool catalogs, file contents, and conversation history
  • 06Wrapping production AI agents with transparent compression that preserves OAuth credentials and subscription logins

Related MCP Servers

View more

public-apis

Featured

A collective list of free APIs

MCP Server461.9k

awesome-python

The definitive list that answers "I want to do X in Python, which tool should I use?"

MCP Server314.3k

awesome-selfhosted

A list of Free Software network services and web applications which can be hosted on your own servers

MCP Server313.1k

ECC

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

MCP Server240.5k

caveman — FAQ

What is the Caveman MCP server?+

Caveman is an MCP server and CLI tool that reduces AI agent token usage through two mechanisms: a skill that makes agents respond concisely in "caveman-speak," and a local proxy that compresses input payloads (files, logs, tool schemas) before they reach the language model provider. Both preserve code accuracy and support byte-exact recovery of original content.

How do I install Caveman for Claude Code?+

Install globally with 'npm install -g @caveman-ai/cli && caveman setup --install' then run 'caveman claude' to wrap Claude Code with the proxy. For the skill only, use 'npx skills add JuliusBrussee/caveman' or 'claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman'.

Which AI clients and agents work with Caveman?+

Caveman has native wrapper support for Claude Code, OpenAI Codex CLI, Gemini CLI, Aider, opencode, Hermes Agent, and OpenClaw. The skill works in 30+ agents including Cursor, Windsurf, Cline, and Copilot. Any framework using baseURL configuration (Vercel AI SDK, LangChain, LiteLLM) can point to the local proxy.

Does Caveman require API keys or a subscription?+

Caveman itself is free and requires no API keys or subscription. It works as a local proxy and skill, forwarding requests to your existing language model provider (Anthropic, OpenAI, Google) using your existing credentials. OAuth tokens for Claude Pro/Max and ChatGPT subscriptions pass through unchanged.

Is Caveman free and open source?+

The skill, CLI, and adoption surfaces are MIT-licensed and free forever. The engine-linked runtime (proxy, compression engine, MCP server) is BSL-1.1 source-available, free for first-party use including production, but requires a commercial license for third-party hosted services. All BSL code converts to Apache-2.0 in 2030 or four years after release.

What are the prerequisites for running Caveman?+

Node.js 18+ is required for installation. The proxy needs Go and pnpm if building from source instead of using signed binaries. Pixel mode and browser compression require Chrome to be installed. The tool runs entirely locally with no external Caveman account needed.

How do I install caveman?+

Open the source repository on GitHub and follow its README. caveman is a mcp server — MCP Agents Market links you directly to the official repo.

Is caveman free?+

caveman is an open-source project hosted on GitHub. Check the repository for its license and any usage requirements.

Related searches

Developer Tools MCP ServersProductivity & Workflow MCP ServersBest SkillsBest PluginsBest AgentsTop rated listingsClaude Code toolsBrowse all categories