@yabasha/cas

A CLI scaffolding tool for the Composable AI Stack — quickly bootstrap a Bun + Turborepo + Convex monorepo for AI-powered applications, with optional components you can opt into.

What you get

The template the CLI clones is a hardened, production-ready monorepo for AI-integrated SaaS apps. Highlights from the latest cut:

• Env validation via @acme/config — a Zod schema with NODE_ENV-conditional required fields and a frozen env export.
• Convex Auth (@convex-dev/auth with Password provider) plus requireUser / requireUserDoc guards.
• Stripe webhook handler at POST /stripe/webhook with async signature verification (Web Crypto), idempotency markers written after successful processing, and Stripe-friendly retries on failure.
• Hardened Elysia gateway with pino structured logging (header redaction), CORS, timingSafeEqual bearer guard, in-memory rate limiter with periodic sweep, and SIGINT/SIGTERM graceful shutdown.
• CSP middleware with per-request nonces, 'strict-dynamic', dev-only 'unsafe-eval', and upgrade-insecure-requests gated to production.
• @acme/ai guardrail pattern (runWithGuardrails) — Vercel AI SDK + Zod input/output schemas + moderation hook + automatic Langfuse tracing.
• BullMQ worker with Convex scheduled functions for cron-like jobs.
• Eval harness (@acme/evals) with Langfuse scoring and non-zero exit on regression.

Layout

• apps/web — Next.js 16 (App Router) + Tailwind 4 + shadcn/ui, CSP middleware
• apps/convex (always) — Convex backend, Convex Auth, Stripe webhook, idempotency table
• apps/api (optional) — Hardened ElysiaJS API gateway on Bun
• apps/worker (optional) — BullMQ background worker & eval runner (Bun)
• packages/ai (always) — Vercel AI SDK + Langfuse tracing + guardrail pattern
• packages/config (always) — Zod env schema + dotenv loader
• packages/prompts (always) — versioned prompt templates
• packages/schemas (always) — Zod schema definitions
• packages/shared (always) — shared utilities
• packages/evals (optional) — eval harness with Langfuse scoring
• tooling/scripts/check-versions.ts — fails CI if pinned versions drift

Why packages/config and packages/ai are always included: they are workspace dependencies of apps/convex (env validation) and apps/worker (LLM calls). Removing them would produce a broken scaffold.

Pinned tech stack (template defaults)

The template ships with bun run check:versions so CI can fail if any pinned dependency drifts.

Features

• Interactive & non-interactive modes — guided prompts or CLI flags
• Component selection — opt into:
  • API Service (apps/api, hardened ElysiaJS on Bun)
  • Background Worker (apps/worker, BullMQ + Convex scheduled functions)
  • AI Evaluations (packages/evals, Langfuse-scored)
• Template customization — replaces project name, author, license, and year across all package.json files
• Multi-package-manager support — bun (default), npm, yarn, pnpm
• Dry-run mode — preview the plan without writing anything
• Git initialization — fresh repo by default (--no-git to skip)
• Update notifications — checks npm for newer versions of @yabasha/cas

Installation

# Global installation with bun (recommended)
bun add -g @yabasha/cas

# Or with npm (Node.js fallback)
npm install -g @yabasha/cas

Usage

Interactive Mode

Run without arguments to use the interactive prompt:

cas init

This will guide you through:

1. Project name
2. Author name
3. License selection
4. Component selection
5. Package manager choice

Non-interactive Mode

Provide all options via CLI flags:

# Create with all optional components
cas init my-project --all

# Create minimal project (no optional components)
cas init my-project --minimal

# Select specific components
cas init my-project --with-api --with-worker

# Full customization
cas init my-project \
  --author "Your Name" \
  --license MIT \
  --package-manager bun \
  --with-api \
  --with-evals

Using bunx/npx

# With bun (recommended)
bunx @yabasha/cas init my-project

# With npm (Node.js fallback)
npx @yabasha/cas init my-project

CLI Options

Supported Licenses

• MIT (default)
• Apache-2.0
• ISC
• GPL-3.0
• BSD-3-Clause
• UNLICENSED

Examples

Create a full-stack AI project

cas init my-ai-app --all --author "Jane Doe" --license MIT

Create API-only project

cas init my-api --with-api --minimal

Preview what would be created

cas init my-project --all --dry-run

Force recreate existing project

cas init my-project --force --all

Skip automatic dependency installation

cas init my-project --all --no-install

Project Structure

After scaffolding with --all, you'll get:

my-project/
├── apps/
│   ├── web/          # Next.js 16 + Tailwind 4 + shadcn/ui + CSP middleware
│   ├── convex/       # Convex + Convex Auth + Stripe webhooks (primary backend)
│   ├── api/          # Hardened Elysia gateway (optional)
│   └── worker/       # BullMQ + Convex scheduled functions (optional)
├── packages/
│   ├── ai/           # Vercel AI SDK + Langfuse tracing + runWithGuardrails
│   ├── config/       # Zod env schema + dotenv loader
│   ├── prompts/      # Versioned prompt templates
│   ├── schemas/      # Zod schema definitions
│   ├── shared/       # Shared utilities
│   └── evals/        # Eval harness with Langfuse scoring (optional)
├── tooling/
│   └── scripts/      # check-versions and other repo tooling
├── .env.example      # Required env surface (Convex, Stripe, LLM, Langfuse, Redis)
├── turbo.json
├── package.json
├── CLAUDE.md         # Guidance for Claude / Claude Code
├── AGENTS.md
└── README.md

Environment variables

After scaffolding, copy .env.example to .env and fill in:

Convex secrets live in the Convex dashboard, not .env. Use bunx convex env set <NAME> <VALUE> from apps/convex.

Scaling the rate limiter (apps/api)

apps/api ships an in-memory rate limiter (src/middleware/rate-limit.ts) that's fine for single-instance deployments. For multi-worker/multi-region deployments, swap the backing store for Upstash or Redis:

1. bun add --filter=api @upstash/ratelimit @upstash/redis
2. Replace the Map in rate-limit.ts with an Upstash Ratelimit instance.
3. Set UPSTASH_REDIS_REST_URL and UPSTASH_REDIS_REST_TOKEN.

The middleware exposes X-RateLimit-Backend: memory so you can verify which store is in use at runtime.

Programmatic API

You can also use the scaffolding functionality programmatically:

import { scaffold, validateProjectName, slugify } from '@yabasha/cas'

// Scaffold a new project
const result = await scaffold({
  projectName: 'my-project',
  dir: './my-project',
  author: 'Your Name',
  license: 'MIT',
  withApi: true,
  withWorker: true,
  withEvals: false,
  withConfig: true, // ignored — packages/config is always scaffolded
  all: false,
  minimal: false,
  force: false,
  noInstall: false,
  noGit: false,
  packageManager: 'bun',
  dryRun: false,
})

if (result.success) {
  console.log('Project created at:', result.dir)
}

// Validate project name (npm conventions)
const validation = validateProjectName('my-project')
if (!validation.valid) {
  console.error(validation.message)
}

// Convert name to slug
const slug = slugify('My Project Name') // 'my-project-name'

Exported Functions

Development

Prerequisites

• Bun (recommended)
• Node.js 20+ (fallback)

Setup

# Clone the repository
git clone https://github.com/yabasha/cas.git
cd cas

# Install dependencies
bun install

# Build the project
bun run build

# Run in development mode
bun run dev

Scripts

Testing

# Run all tests
bun test

# Watch mode
bun run test:watch

Project Structure

src/
├── __tests__/        # Unit tests
├── args.ts           # CLI argument parsing
├── banner.ts         # ASCII banner display
├── cli.ts            # Main CLI entry point
├── index.ts          # Public API exports
├── prompts.ts        # Interactive prompts
├── template.ts       # Template processing logic
├── types.ts          # TypeScript type definitions
└── utils.ts          # Utility functions

Requirements

• Bun (recommended) or Node.js 20+ (fallback)
• Internet connection (for cloning the template repository)
• Git (for repository initialization)

Note: This CLI is built with full Bun support. Node.js is supported as a fallback runtime and will be automatically detected.

Related

• Composable AI Stack - The template repository

License

MIT