🚀 AI Stack Orchestrator v2

Config-driven macOS AI development environment — extensible, reliable, and maintainable.

Add a new agent = add a folder. No code changes.

---

Architecture

AI_SETUP/
├── setup.sh                  # Bash bootstrap → installs brew+node, runs orchestrator
├── nuke.sh                   # Bash bootstrap → runs teardown
├── package.json / tsconfig.json
│
├── src/
│   ├── orchestrate.ts        # Main entry: foundations → select agents → install → MCP → auth
│   ├── nuke.ts               # Teardown: select agents → uninstall → cleanup → remove foundations
│   ├── ai-init.ts            # Project scaffolder: select agent → copy scaffold + plugins
│   ├── core/
│   │   ├── types.ts          # AgentConfig, AgentDriver, PluginConfig, Context interfaces
│   │   ├── registry.ts       # Auto-discovers agents/ and plugins/ folders
│   │   ├── config.ts         # Reads + validates agent.json
│   │   └── runner.ts         # Builds Context objects for drivers
│   ├── lib/
│   │   ├── ui.ts             # Terminal output (chalk colors, headers, steps)
│   │   ├── shell.ts          # exec, retry, safeRm, sudoRm
│   │   ├── prompts.ts        # Interactive menus (inquirer)
│   │   ├── mcp.ts            # Write MCP configs to disk
│   │   └── template.ts       # Deep-copy scaffold dirs with {{VAR}} replacement
│   └── foundations/
│       ├── index.ts           # Install/uninstall in order: brew → node → python → git
│       ├── homebrew.ts
│       ├── node.ts
│       ├── python.ts
│       └── git.ts
│
├── agents/                    # One folder per agent (data, not code)
│   ├── claude/
│   │   ├── agent.json         # Config: name, MCP paths, cleanup paths
│   │   ├── driver.ts          # 4 functions: isInstalled, install, authenticate, uninstall
│   │   ├── mcp/               # Global MCP JSON files
│   │   └── scaffold/          # Files copied into new projects by ai-init
│   ├── codex/
│   │   ├── agent.json         # Config: name, MCP paths (TOML), cleanup paths
│   │   ├── driver.ts          # npm install, ChatGPT auth, npm uninstall
│   │   ├── mcp/               # Global MCP TOML (merged into ~/.codex/config.toml)
│   │   └── scaffold/          # AGENTS.md, .codex/, .agents/skills/
│   ├── cursor/
│   ├── vscode/
│   └── _template/             # Copy this to add a new agent
│
└── plugins/                   # Optional cross-agent project add-ons
    └── _template/

---

Quick Start

# Setup everything
chmod +x setup.sh && ./setup.sh

# Scaffold a new project
ai-init

# Teardown everything
chmod +x nuke.sh && ./nuke.sh

# Dry run (see what would happen without doing it)
./setup.sh --dry-run
./nuke.sh --dry-run

---

Built-in Agents

Agent	Icon	Global MCPs	Project Scaffold
Claude Code	🤖	sequential-thinking, memory	CLAUDE.md, rules, skills, .mcp.json
VS Code	💎	puppeteer	copilot-instructions.md, .instructions.md, skills, .vscode/mcp.json

---

Adding a New Agent

# 1. Copy the template
cp -r agents/_template agents/opencode

# 2. Edit the config
#    agents/opencode/agent.json — name, MCP paths, cleanup paths
#    agents/opencode/driver.ts  — isInstalled, install, authenticate, uninstall

# 3. Add MCP configs
#    agents/opencode/mcp/global.json

# 4. Add scaffold files
#    agents/opencode/scaffold/  — rules, skills, project MCP, etc.

# 5. Done. Run ./setup.sh — OpenCode appears in the menu automatically.

agent.json Reference

{
  "name": "OpenCode",              // Display name
  "id": "opencode",               // Must match folder name
  "description": "Terminal AI",    // Shown in menus
  "icon": "🖥️",                   // Menu icon
  "requires": ["homebrew", "node"],
  "installMethod": "custom",      // "custom" | "brew-cask" | "brew" | "npm-global"
  "command": "opencode",           // CLI command to check installation
  "mcp": {
    "global.json": {               // File under mcp/ → destination path
      "path": "~/.opencode/mcp.json"
    }
  },
  "cleanup": {
    "paths": ["~/.opencode"],      // Removed by nuke (user-level)
    "sudoPaths": []                // Removed by nuke (with sudo)
  }
}

driver.ts Reference

import type { AgentDriver, Context } from "../../src/core/types.js";

export const driver: AgentDriver = {
  async isInstalled() { /* return true if installed */ },
  async install(ctx)  { /* install the agent */ },
  async authenticate(ctx) { /* prompt user to log in */ },
  async uninstall(ctx) { /* remove the agent */ },
};
export default driver;

---

Codex Agent — Deep Dive

Codex is unique among the agents because it uses TOML configuration (not JSON) and has a richer per-project structure: subagents, command rules, skills, and shared .agents/ skills.

Global Setup (./setup.sh → select Codex)

Installs the CLI via npm install -g @openai/codex, then merges MCP servers into ~/.codex/config.toml (preserving your existing model, sandbox, and approval settings):

# Your existing settings stay untouched
# model = "gpt-5.4"
# approval_policy = "on-request"

[mcp_servers.context7]
command = "npx"
args = ["-y", "@upstash/context7-mcp"]

[mcp_servers.sequential-thinking]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-sequential-thinking"]

[mcp_servers.memory]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-memory"]

Project Scaffold (ai-init → select Codex → "myapp")

myapp/
├── AGENTS.md                              # Project instructions for Codex
├── .agents/
│   └── skills/                            # Shared skills (usable by any subagent)
│       ├── code-review/
│       │   └── SKILL.md                   # Structured review workflow
│       └── testing-strategy/
│           └── SKILL.md                   # Unit/integration/E2E patterns
└── .codex/
    ├── config.toml                        # Project-level MCP servers + settings
    ├── agents/
    │   ├── explorer.toml                  # Read-only codebase exploration (skills disabled)
    │   └── reviewer.toml                  # PR review (code-review + testing-strategy enabled)
    ├── rules/
    │   └── default.rules                  # Starlark command policies
    └── skills/
        └── project-conventions/
            └── SKILL.md                   # Coding standards, git workflow, docs

TOML Support

The framework detects config format from file extension:

Source file	Read as	Destination file	Write as
mcp/global.json	JSON	~/.cursor/mcp.json	JSON (overwrite)
mcp/global.toml	TOML	~/.codex/config.toml	TOML (merge)

TOML destinations are merged (not overwritten) — your existing config.toml settings are preserved. JSON destinations use the existing overwrite behavior.

Subagents

Each .toml file in .codex/agents/ defines a subagent with its own model, sandbox, and skill configuration:

# .codex/agents/reviewer.toml
name = "reviewer"
description = "PR reviewer focused on correctness, security, and missing tests."
sandbox_mode = "read-only"
model_reasoning_effort = "high"

[[skills.config]]
path = ".agents/skills/code-review/SKILL.md"
enabled = true

Subagents toggle shared skills independently — the explorer disables review skills, the reviewer enables them.

Command Rules

.codex/rules/default.rules uses Starlark syntax:

prefix_rule(
    pattern = ["git", ["status", "log", "diff"]],
    decision = "allow",
)
prefix_rule(
    pattern = ["git", ["push", "reset", "rebase"]],
    decision = "prompt",
    justification = "Destructive git operations need human review.",
)

---

Adding a Plugin

Plugins are optional add-ons that inject files into projects during ai-init.

# 1. Copy the template
cp -r plugins/_template plugins/git-hooks

# 2. Edit plugins/git-hooks/plugin.json
{
  "name": "Git Hooks",
  "id": "git-hooks",
  "description": "Husky + commitlint pre-commit hooks",
  "icon": "🪝",
  "compatibleAgents": []    // empty = compatible with all agents
}

# 3. Add files under plugins/git-hooks/files/
#    These are copied to the project root preserving directory structure.
#    Supports {{PROJECT_DIR}} and {{PROJECT_NAME}} tokens.

# 4. Done. Plugin appears in ai-init menu automatically.

---

How It Works

Setup Flow (./setup.sh)

1. Bash bootstrap ensures Homebrew + Node exist
2. orchestrate.ts installs foundations (brew, node, python, git)
3. Auto-discovers agents/*/agent.json
4. Interactive checkbox: pick which agents to set up
5. For each agent: install → write global MCP → authenticate
6. Injects ai-init shell function into ~/.zshrc

Teardown Flow (./nuke.sh)

1. Safety gate (type NUKE to confirm)
2. Auto-discovers all agents
3. For each: run driver.uninstall() → remove cleanup paths → remove MCP configs
4. Removes ai-init from shell configs
5. Uninstalls foundations in reverse order

Scaffold Flow (ai-init)

1. Prompt for project name
2. Select one agent from discovered list
3. Optionally select plugins
4. Deep-copy agents/<id>/scaffold/ into new project
5. Replace {{PROJECT_DIR}} and {{PROJECT_NAME}} tokens in copied files
6. Copy plugin files into project

---

Design Principles

Principle	Implementation
Convention over configuration	Agents discovered by folder existence — no central registry
Data, not code	MCP configs, rules, skills are plain files — edit with any editor
Typed driver contract	AgentDriver interface enforces consistent 4-function API
Idempotent	Every step checks before acting (already installed → skip)
Dry run	--dry-run flag prints actions without executing
Isolated agents	One agent's broken config can't affect another
Template tokens	{{PROJECT_DIR}} replaced at scaffold time for dynamic values

---

What Each Agent Gets

MCP Servers

Agent	Global MCP(s)	Project MCPs
Claude Code	sequential-thinking + memory	filesystem + github
Codex	context7 + sequential-thinking + memory	git + filesystem
Cursor	fetch	brave-search + git
VS Code	puppeteer	sqlite + postgres

Skills (per project)

#	Claude Code	Codex	Cursor	VS Code
1	api-design — REST, Zod, rate limiting, OpenAPI	code-review — Structured review, security, correctness	react-components — Functional, hooks, lazy loading	typescript-strict — Strict config, generics, utility types
2	error-handling — Error classes, tracing, circuit breakers	testing-strategy — Unit, integration, E2E patterns	tailwind-patterns — Utility-first, responsive, dark mode	full-stack-patterns — Shared types, typed API, e2e safety
3		project-conventions — Code org, git workflow, docs