Which AI Coding Agents Are Supported by Caveman? A Complete Compatibility Guide

Caveman supports 30+ AI coding agents out-of-the-box, including Claude Code, Gemini CLI, opencode, Cursor, and GitHub Copilot, using either automatic hook activation or static rule files depending on the agent's extension capabilities.

Caveman is a compression skill that attaches to AI coding agents to reduce token usage by replacing verbose output with terse "caveman-speak" while preserving code and error messages byte-for-byte. According to the JuliusBrussee/caveman repository, the installer detects and configures agents through two primary mechanisms: JavaScript hooks for agents with plugin systems, and static markdown rule files for agents that read configuration directories.

Agents with Full Auto-Activation Support

The following AI coding agents support automatic activation via hooks or plugins that fire when a session starts. These agents require no manual intervention after initial installation.

Claude Code

Claude Code offers the most seamless integration. The installer adds the skill via the plugin marketplace and activates it automatically on every session start.

# Install via marketplace

claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman

Once installed, the hook file src/hooks/caveman-activate.js (copied to ~/.claude/) automatically enables compression mode. The agent also loads src/hooks/caveman-stats.js to report token savings to the status line.

Gemini CLI

Gemini CLI supports automatic activation through its extension system.

gemini extensions install https://github.com/JuliusBrussee/caveman

The extension activates immediately upon session start without requiring manual commands.

opencode

opencode receives a full plugin installation including sub-agents. The installer adds a plugin directory under ~/.config/opencode/plugins/caveman/ and injects Caveman agents into ~/.config/opencode/agents/AGENTS.md.

node bin/install.js --only opencode

# Or via npx

npx -y github:JuliusBrussee/caveman -- --only opencode

This installation includes the cavecrew sub-agents located in agents/cavecrew-reviewer.md, allowing opencode-based agents to benefit from Caveman's compression.

OpenClaw and Hermes Agent

Both OpenClaw and Hermes Agent support auto-activation through npx-based installation:

# OpenClaw (adds workspace skill and marker block to SOUL.md)

npx -y github:JuliusBrussee/caveman -- --only openclaw

# Hermes Agent

npx -y github:JuliusBrussee/caveman -- --only hermes

Agents Requiring Manual Activation or Static Rules

For agents without hook systems, Caveman drops a static rule file (caveman-activate.md) into the agent's per-repo configuration directory. These installations require either a manual /caveman command each session or the --with-init flag for permanent rule installation.

Codex CLI

Codex CLI requires manual activation per session despite having plugin support:

npx skills add JuliusBrussee/caveman -a codex

After installation, you must type /caveman at the start of each session to activate compression mode.

Cursor, Windsurf, and Cline

These popular VS Code-based agents lack hook systems but support static rule files:

# Install for Cursor (manual per-session activation)

npx skills add JuliusBrussee/caveman -a cursor

# Install for Cursor with permanent rule file

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

Using --with-init creates ./.cursor/rules/caveman.mdc (or equivalent paths for Windsurf and Cline), making the /caveman command available without reinstalling. Without this flag, you must run the install command each session.

GitHub Copilot

GitHub Copilot receives "soft probe" support through static rules:

npx -y github:JuliusBrussee/caveman -- --only copilot --with-init

This drops the activation rules into Copilot's configuration directory, though you must still manually trigger compression with /caveman commands.

Continue, Kilo Code, Roo Code, and Augment Code

These agents and others in the "30+" category use profile-based installation:

npx skills add JuliusBrussee/caveman -a <profile>

Replace <profile> with continue, kilo-code, roo-code, or augment-code. To see the full matrix of supported agents, run:

node bin/install.js --list

How Caveman Integrates with Different Agent Architectures

The installer in bin/install.js contains a PROVIDERS array that defines the full compatibility matrix. It detects which agents are present on the host and routes installation through three architectural paths:

1. Skill Files with Hooks For agents like Claude Code, the installer writes JavaScript hooks (caveman-activate.js, caveman-mode-tracker.js) into the agent's config directory. These hooks replace verbose output fragments with compressed "caveman-speak" automatically when sessions start.

2. Static Rule Files For agents without extension APIs (Cursor, Windsurf, Cline), the installer copies src/rules/caveman-activate.md into the agent's per-repo rules directory. This file contains the /caveman, /caveman-commit, and other command definitions, plus the "caveman-begin / caveman-end" markers that trigger compression.

3. Plugin Systems For opencode and similar agents, the installer creates a full plugin structure under src/plugins/opencode/commands/caveman.md, adding the Caveman commands to the agent's native command palette.

All installation methods write a flag file (.caveman-active) that the runtime checks to determine the active compression level.

Practical Installation Examples

Quick Start with Claude Code

The default installation method targets Claude Code with automatic activation:

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

Start a Claude Code session and type /caveman. The agent responds in compact form, for example: "New ref each render. Wrap in useMemo."

Permanent Setup for Cursor

To enable Caveman across all Cursor projects without reinstalling:

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

This creates .cursor/rules/caveman.mdc in your project root. In any subsequent session, trigger full compression with:

/caveman full

Workspace Integration for opencode

Add Caveman to an opencode workspace including sub-agents:

node bin/install.js --only opencode

After installation, invoke sub-agents like the reviewer:

/caveman-review

This returns a one-line PR comment from the agents/cavecrew-reviewer.md sub-agent.

Summary

• Caveman supports 30+ AI coding agents through the installer in bin/install.js, which detects available agents via the PROVIDERS array.
• Auto-activation works with Claude Code, Gemini CLI, opencode, OpenClaw, and Hermes Agent using JavaScript hooks or native plugins.
• Manual activation is required for Codex CLI, Cursor, Windsurf, Cline, GitHub Copilot, Continue, Kilo Code, Roo Code, and Augment Code.
• Permanent installation using --with-init drops static rule files (caveman-activate.md) into agent configuration directories, eliminating the need for per-session setup.
• Core files include src/hooks/caveman-activate.js for hooked agents and src/rules/caveman-activate.md for rule-based agents.

Frequently Asked Questions

Does Caveman work with VS Code extensions like GitHub Copilot?

Yes, Caveman supports GitHub Copilot and other VS Code-based agents through static rule files. While these agents lack hook systems for automatic activation, running node bin/install.js --only copilot --with-init drops the activation rules into the appropriate configuration directory, allowing you to trigger compression manually with /caveman commands.

How do I check if my AI coding agent is compatible with Caveman?

Run node bin/install.js --list from the Caveman repository to see the full matrix of supported agents. The installer auto-detects which agents are present on your system and can configure multiple agents simultaneously. If your agent supports reading static rule files or has a plugin/extension system, it is compatible.

What is the difference between --only and --with-init flags?

The --only flag restricts installation to a specific agent (e.g., --only cursor), while --with-init creates a permanent rule file in the agent's configuration directory rather than requiring you to run the install command each session. Use both together (e.g., --only cursor --with-init) for agents like Cursor that lack auto-activation hooks.

Can I use Caveman with multiple agents on the same machine?

Yes. The installer can detect and configure multiple agents simultaneously. You can run the global install script to set up all detected agents, or use specific flags to install Caveman for particular agents while maintaining separate configurations for each environment.

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