Download the PHP package thinkliveid/crew without Composer

Crew

A PHP CLI tool for managing AI agent skills, sub-agents, and slash commands across multiple IDE coding assistants.

Crew discovers which AI agents are installed on your system, downloads resources from GitHub repositories, syncs local definitions, and writes configurations to the correct agent-specific directories — all from a single command.

Requirements

• PHP 8.3+
• Composer

Installation

Quick Start

Supported Agents

Crew auto-detects the following AI coding assistants on your system and in your project:

When you run crew install:skill, only detected agents are shown for selection. Resources are written to the correct paths for each selected agent.

Commands

Crew organises commands into four groups across three resource types, plus top-level init and sync commands that run installers/updaters in one pass:

init — Install everything at once

Runs install:skill and install:subagent in sequence. The first step performs agent detection and saves the selection to crew.json; the second reuses that selection, so you're only prompted for agents once.

sync — Update everything at once

Runs update:skill, update:subagent, and update:command in sequence. Each step reuses the agents saved in crew.json and refreshes local resources plus any configured GitHub sources.

new:* — Scaffold a new resource locally

Create a new skill, sub-agent, or slash command definition from scratch with an interactive prompt.

Generated files:

Names must be lowercase alphanumeric with hyphens (1–64 chars, no leading/trailing/consecutive hyphens). new:subagent also prompts for a model choice (sonnet, opus, haiku, or inherit).

install:* — Full setup flow

Runs the complete setup: detect agents, select them, sync local resources, and install from GitHub.

Non-interactive mode (uses auto-detected agents, skips prompts):

add:* — Add from a GitHub repository

Discover and install resources from a GitHub repository. The repository is saved to crew.json for future installs and updates.

update:* — Update to the latest version

Re-runs the install flow in non-interactive mode to refresh all local and GitHub resources.

Configuration

Crew stores its configuration in crew.json at the project root:

GitHub Authentication

For private repositories or to avoid API rate limits, set a GitHub token:

Or use an environment variable:

Project Structure

Local Resources

Place resources in your project's .ai/ directory:

Running crew install:skill copies skills from .ai/skills/ to each selected agent's skill directory (e.g., .claude/skills/my-skill/, .junie/skills/my-skill/). The same pattern applies to sub-agents and slash commands.

Remote Resources (GitHub)

Crew discovers resources in GitHub repositories by traversing the repo tree and looking for marker files (SKILL.md, agent markdown files, slash-command markdown files).

Architecture

Crew is built with a contract-driven architecture. Each agent implements interfaces that define its capabilities:

• SupportsSkills — agent can receive skill files
• SupportsSubAgents — agent can receive sub-agent definitions
• SupportsCommands — agent can receive slash command files
• SupportsGuidelines — agent can receive guideline files
• SupportsMcp — agent can receive MCP server configurations

The detection system uses a strategy pattern to discover agents:

• DirectoryDetectionStrategy — checks for directories/paths (supports glob patterns)
• FileDetectionStrategy — checks for specific files in the project
• CommandDetectionStrategy — runs shell commands to detect installed tools
• CompositeDetectionStrategy — combines multiple strategies with OR logic

Running in a Container (Claude on Host)

If crew runs inside a container but the AI agent (e.g. Claude Code) is installed on the host, the container's command -v claude won't see the host binary. The fix is to bind-mount the host's binary and config into the container so command -v claude resolves and crew's system detection passes.

Find the host paths first

docker-compose.yml

docker run one-liner

.devcontainer/devcontainer.json

Notes

• The container must have Node.js available — claude is a Node CLI. Use a base image with Node, or install it in your Dockerfile.
• Mount targets must match the container user's $HOME: /root/... for root images, /home/vscode/... for the Microsoft devcontainer images.
• macOS quirk: when you log in via claude /login on a macOS host, credentials may live in the Keychain rather than ~/.claude/.credentials.json. Bind-mounting ~/.claude won't carry those credentials into a Linux container — set ANTHROPIC_API_KEY in the container env, or run claude /login once from inside the container so credentials land in the mounted file.
• Once the mounts are active, command -v claude succeeds inside the container and crew install:skill detects claude_code with no code changes.

If you can't bind-mount the binary, you can still tell crew which agents are present by listing them in crew.json and passing --skip-detection:

License

MIT