Agent-Specific Installation for Caveman: Targeted Setup Guide for AI Agents

You can install the Caveman skill for a specific AI agent using the --only <id> flag with the installer script, or target individual agents like Claude Code, Gemini CLI, Opencode, and Cursor using their native CLI commands.

The Caveman skill compresses AI agent output by removing filler words while preserving technical accuracy. According to the JuliusBrussee/caveman source code, the installation is handled by a single Node.js script (bin/install.js) that detects installed agents and executes the appropriate native commands for each platform. You can deploy globally across all detected agents or isolate the installation to a single agent using specific flags.

Global Installation vs. Agent-Specific Installation

The installer supports two installation modes. Global installation detects every supported agent on your PATH and configures Caveman for all of them simultaneously. Agent-specific installation targets only one agent using the --only flag, which is useful when you want to test the skill on a single platform or when working with agents that require different configuration approaches.

The core logic resides in bin/install.js, which references a PROVIDERS matrix mapping agent IDs to their specific CLI installation commands.

Installing for All Agents (One-Liner)

To install Caveman globally across every supported agent on your system, run the appropriate one-liner for your shell.

macOS, Linux, or WSL:

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash

Windows PowerShell 5.1+:

irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex

This script completes three stages: it detects which agents are installed (Claude Code, Gemini CLI, Opencode, OpenClaw, etc.), runs each agent's native install command, and wires up hook files and the .caveman-active flag file that stores your selected compression level. The process typically completes in approximately 30 seconds and is idempotent—safe to run multiple times.

Targeting a Specific Agent

When you only want Caveman active for one agent, use the --only <id> flag to bypass global detection. The installer supports the following agent IDs and commands as defined in the PROVIDERS matrix in bin/install.js:

Agent	ID	Native Install Command
Claude Code	claude	claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman
Gemini CLI	gemini	gemini extensions install https://github.com/JuliusBrussee/caveman
Opencode	opencode	npx -y github:JuliusBrussee/caveman -- --only opencode
OpenClaw	openclaw	npx -y github:JuliusBrussee/caveman -- --only openclaw
Cursor	cursor	npx skills add JuliusBrussee/caveman -a cursor

To install only for Claude Code via the shell script:

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash -s -- --only claude

To install only for Cursor using the local Node installer:

node bin/install.js --only cursor

Note that Cursor and similar editors do not support automatic activation hooks, requiring manual /caveman commands each session unless you deploy the always-on rule files described below.

Always-On Rule Files for Agents Without Hooks

Some agents—including Cursor, Windsurf, Cline, and Copilot—lack a hook system for automatic activation. For these platforms, Caveman writes a static rule file (caveman-activate.md) to the agent's configuration directory, making the skill active by default without requiring manual commands.

To manually place the rule file for Cursor:

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/src/rules/caveman-activate.md \
  > .cursor/rules/caveman.mdc

To generate rule files for multiple agents simultaneously, use the --with-init flag:

node bin/install.js --with-init --only cursor --only windsurf

This writes the contents of src/rules/caveman-activate.md to each specified agent's rules directory, ensuring Caveman activates automatically when the agent loads the project.

Verifying Your Installation

After running the installer, confirm the skill is properly configured using the following methods.

List detected agents:

node bin/install.js --list

This outputs a table with checkmarks (✓) indicating which agents have Caveman installed.

Check the activation flag (Claude Code only):

cat "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/.caveman-active"

Expected output: full (or the compression level you selected during install).

Test the skill: Open your targeted agent and issue the /caveman command or ask it to "talk like caveman." The response should be terse, omitting articles and filler words while retaining technical precision.

Uninstalling Caveman from Specific Agents

To remove Caveman entirely, including hooks and plugin registrations, run the uninstall command:

npx -y github:JuliusBrussee/caveman -- --uninstall

This command removes hook entries from Claude Code's settings.json, deletes generated hook scripts (caveman-*.js, caveman-statusline.*), uninstalls per-agent plugins, removes the Opencode plugin directory, strips the OpenClaw workspace block, and deletes the .caveman-active flag file.

Static rule files created with --with-init are not removed automatically. You must manually delete these files from your agent's rules directory if you wish to completely remove the skill.

Summary

• Global install uses the one-liner shell script to configure all detected agents simultaneously.
• Agent-specific install uses --only <id> with valid IDs: claude, gemini, opencode, openclaw, cursor, etc.
• Hook-based agents (Claude Code, Gemini CLI) receive automatic activation via the .caveman-active flag and native plugin hooks.
• Rule-based agents (Cursor, Windsurf) require caveman-activate.md files deployed via --with-init for automatic activation.
• Core files: bin/install.js contains the installer logic, src/rules/caveman-activate.md provides the always-on rules, and INSTALL.md documents the full provider matrix.

Frequently Asked Questions

How do I install Caveman only for Claude Code without affecting other agents?

Use the --only claude flag with the installation script. Run: curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash -s -- --only claude. This executes only the claude plugin marketplace add and claude plugin install commands, skipping detection for Gemini, Opencode, or other agents.

Why does Cursor require different installation steps than Claude Code?

Cursor does not expose a hook system for automatic skill activation, according to the PROVIDERS matrix in bin/install.js. While Claude Code supports hooks and a .caveman-active flag file, Cursor requires either the npx skills add command for manual /caveman activation, or a static rule file (caveman-activate.md) placed in the .cursor/rules/ directory for always-on behavior.

Where does the installer store the configuration for active agents?

For Claude Code, the installer writes a .caveman-active flag file to $CLAUDE_CONFIG_DIR (or $HOME/.claude if the variable is unset), containing the compression level (e.g., full). For agents without hook support, the installer writes caveman-activate.md to agent-specific rules directories within the project folder, such as .cursor/rules/caveman.mdc.

Can I safely run the installer multiple times for different agents?

Yes. The installer is idempotent and safe to re-run. Each execution with --only <id> adds Caveman only to the specified agent without removing existing installations from other agents. To remove Caveman from a specific agent, you must use the --uninstall flag, which cleans up hooks, plugins, and flag files across all configured agents.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}