README

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

composer require thinkliveid/crew

Quick Start

# One-shot setup: install skills and sub-agents together
crew init

# Refresh everything from saved config in one pass
crew sync

# Or run a single resource type
crew install:skill

# Create a new skill from scratch
crew new:skill

# Add skills from a GitHub repository
crew add:skill owner/repo

# Update all configured skills to the latest version
crew update:skill

Supported Agents

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

Agent	Skills Path	MCP Config Path
Claude Code	`.claude/skills`	`.mcp.json`
Cursor	`.cursor/skills`	`.cursor/mcp.json`
GitHub Copilot	`.github/skills`	`.vscode/mcp.json`
Gemini CLI	`.agents/skills`	`.gemini/settings.json`
Junie	`.junie/skills`	`.junie/mcp/mcp.json`
Codex	`.agents/skills`	`.codex/config.toml`
OpenCode	`.agents/skills`	`opencode.json`

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:

	Skills	Sub-agents	Commands
new	`new:skill`	`new:subagent`	`new:command`
install	`install:skill`	`install:subagent`	`install:command`
add	`add:skill`	`add:subagent`	`add:command`
update	`update:skill`	`update:subagent`	`update:command`

`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.

crew init

# Non-interactive: uses saved/auto-detected agents, skips all prompts
crew init --no-interaction

`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.

crew sync

`new:*` — Scaffold a new resource locally

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

# Interactive — prompts for name and description
crew new:skill
crew new:subagent
crew new:command

# Provide the name upfront to skip the name prompt
crew new:skill my-skill
crew new:subagent my-agent
crew new:command my-command

Generated files:

Command	Output
`new:skill`	`.ai/skills/{name}/SKILL.md`
`new:subagent`	`.ai/agents/{name}.md`
`new:command`	`.ai/commands/{name}.md`

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.

crew install:skill
crew install:subagent
crew install:command

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

crew install:skill --no-interaction

`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.

# Interactive — prompts for repository if not provided
crew add:skill
crew add:subagent
crew add:command

# Direct — provide the repository
crew add:skill owner/repo
crew add:subagent owner/repo
crew add:command owner/repo

# Full GitHub URL with branch and subdirectory
crew add:skill https://github.com/owner/repo/tree/main/path/to/skills

`update:*` — Update to the latest version

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

crew update:skill
crew update:subagent
crew update:command

Configuration

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

{
    "agents": ["claude_code", "junie"],
    "skills": ["owner/repo"],
    "subagents": ["owner/repo"],
    "commands": ["owner/repo"]
}

Key	Type	Description
`agents`	`string[]`	Selected agent identifiers
`skills`	`string[]`	GitHub repositories to install skills from
`subagents`	`string[]`	GitHub repositories to install sub-agents from
`commands`	`string[]`	GitHub repositories to install slash commands from
`guidelines`	`bool`	Whether guideline writing is enabled
`mcp`	`bool`	Whether MCP configuration is enabled
`github_token`	`string`	GitHub API token for private repositories

GitHub Authentication

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

{
    "github_token": "ghp_your_token_here"
}

Or use an environment variable:

export GITHUB_TOKEN=ghp_your_token_here

Project Structure

Local Resources

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

.ai/
  skills/
    my-skill/
      SKILL.md
      other-files...
  agents/
    my-agent.md
eq  commands/
    my-command.md

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

which claude          # e.g. /Users/you/.npm-global/bin/claude
ls ~/.claude          # config, settings, credentials

docker-compose.yml

services:
  app:
    image: php:8.3-cli
    working_dir: /workspace
    volumes:
      - .:/workspace
      # Claude config (settings, session, credentials)
      - ${HOME}/.claude:/root/.claude
      # Binary into a directory on container PATH
      - ${HOME}/.npm-global/bin/claude:/usr/local/bin/claude:ro
      # If installed via npm, mount its node_modules so the wrapper resolves
      - ${HOME}/.npm-global/lib/node_modules:/usr/local/lib/node_modules:ro
    environment:
      # Optional: macOS Keychain auth doesn't survive into a Linux container.
      # Use an API key, or run `claude /login` once inside the container.
      - ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY}

`docker run` one-liner

docker run --rm -it \
  -v "$PWD:/workspace" -w /workspace \
  -v "$HOME/.claude:/root/.claude" \
  -v "$(which claude):/usr/local/bin/claude:ro" \
  -v "$HOME/.npm-global/lib/node_modules:/usr/local/lib/node_modules:ro" \
  php:8.3-cli bash

`.devcontainer/devcontainer.json`

{
  "image": "mcr.microsoft.com/devcontainers/php:8.3",
  "mounts": [
    "source=${localEnv:HOME}/.claude,target=/home/vscode/.claude,type=bind",
    "source=${localEnv:HOME}/.npm-global/bin/claude,target=/usr/local/bin/claude,type=bind,readonly",
    "source=${localEnv:HOME}/.npm-global/lib/node_modules,target=/usr/local/lib/node_modules,type=bind,readonly"
  ],
  "containerEnv": {
    "ANTHROPIC_API_KEY": "${localEnv:ANTHROPIC_API_KEY}"
  }
}

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:

crew install:skill --skip-detection

License

MIT

thinkliveid / crew

Maintainers

Package info

Statistics

Security

README

Requirements

Installation

Quick Start

Supported Agents

Commands

`init` — Install everything at once

`sync` — Update everything at once

`new:*` — Scaffold a new resource locally

`install:*` — Full setup flow

`add:*` — Add from a GitHub repository

`update:*` — Update to the latest version

Configuration

GitHub Authentication

Project Structure

Local Resources

Remote Resources (GitHub)

Architecture

Running in a Container (Claude on Host)

Find the host paths first

docker-compose.yml

`docker run` one-liner

`.devcontainer/devcontainer.json`

Notes

License

thinkliveid / crew

Maintainers

Package info

Statistics

Security

README

Requirements

Installation

Quick Start

Supported Agents

Commands

init — Install everything at once

sync — Update everything at once

new:* — Scaffold a new resource locally

install:* — Full setup flow

add:* — Add from a GitHub repository

update:* — Update to the latest version

Configuration

GitHub Authentication

Project Structure

Local Resources

Remote Resources (GitHub)

Architecture

Running in a Container (Claude on Host)

Find the host paths first

docker-compose.yml

docker run one-liner

.devcontainer/devcontainer.json

Notes

License

`init` — Install everything at once

`sync` — Update everything at once

`new:*` — Scaffold a new resource locally

`install:*` — Full setup flow

`add:*` — Add from a GitHub repository

`update:*` — Update to the latest version

`docker run` one-liner

`.devcontainer/devcontainer.json`