AI App Starter: Ship a Production AI Assistant SaaS in a Weekend

3–6 weeks of plumbing, done. The parts AI agents usually get subtly wrong ship here solved and tested.

Why not just ask your AI agent to build it?

Claude Code, Cursor, Codex or Antigravity will scaffold a chat app with a pricing page in an hour. The happy path works. Then come the days spent on what it gets subtly wrong, usually discovered in production:

• Payment webhooks that aren't idempotent, so a redelivered event upgrades or credits a user twice.
• Quota race conditions, so parallel requests slip past the daily limit.
• Streaming that breaks on disconnect, or a retry that charges twice.
• Fallbacks that retry the wrong errors: retrying a bad request, or switching models halfway through an answer.
• Razorpay India go-live traps: website approval per domain, UPI Autopay vs card mandates, trial fees as add-ons, total_count deciding the end date.
• Stripe plumbing: Checkout for subscriptions and top-ups, the Customer Portal, signed webhooks, fields that moved between API versions.

This starter ships each of these solved, with tests that prove it, for example test_parallel_requests_cannot_exceed_cap (12 parallel sends against a cap of 5 give exactly 5), test_duplicate_event_id_is_applied_once, test_event_is_recorded_only_after_commit, test_parallel_sends_with_same_key_start_one_run, test_mid_stream_failure_does_not_fall_back and test_second_trial_on_same_upi_id_is_refused.

What this saves you

• Time: about 1–2 weeks of agent iterate-debug-review to reach this tested state (3–6 weeks by hand). By part: auth, streaming and fallback 2–3 days; limits and top-ups 1–2 days; Stripe + Razorpay with idempotent webhooks 2–4 days; memory, admin and deploy 1–2 days. Assumes one experienced developer.
• Money: at ₹1,500–4,000 an hour, that is ₹60,000–3,20,000 of developer time (₹1,80,000–9,60,000 by hand), plus agent token spend. One double-charge or revenue-leak incident can cost ₹20,000–1,00,000 or more, and customer trust.
• Payback: for ₹2,999, it pays for itself in the first afternoon.

Built to be extended by your AI agent

AGENTS.md gives your agent the file map, the invariants it must not break, the test commands and step-by-step recipes (add a model provider, a plan, a top-up pack, an endpoint, a webhook event, a schema change). A CLAUDE.md, a Cursor rule and a Claude Code skill point to it. Your agent extends a tested codebase instead of inventing payment logic.

What's included

• Streaming chat: SSE with stop, regenerate, "explain simpler", conversations, resume after a dropped connection, idempotency keys.
• LLM gateway: Anthropic by default (with Anthropic's server-side refusal fallback), OpenAI or any OpenAI-compatible endpoint, and a mock model; ordered fallbacks, retries, timeouts, token accounting per run.
• Usage limits: free / pro / team plans with daily and monthly caps, top-up packs, free follow-ups, free first retry, burst limits, clear 429 and 402 errors.
• Payments: Stripe (Checkout, Customer Portal, webhooks) and Razorpay (orders, subscriptions with a trial fee, webhooks, trial dedupe, reconcile script), plus a dev simulator so you can test without keys.
• Auth: email + password with argon2id, short-lived JWTs, rotating refresh tokens with reuse detection, optional Google sign-in.
• Memory: notes the assistant learns, retrieved by keyword and recency, and a "what I got wrong" spaced-review example.
• Admin: users, plans, usage by day and model, payments, refunds with a reason saved in the notes, audit log.
• Frontend: Next.js + Tailwind pages for auth, chat, usage meter, pricing with checkout, settings and admin. Mobile-first and accessible.
• Deploy: docker-compose, a single-VPS guide with Caddy and a migrate-then-swap updater, and notes for Render, Fly.io and Railway.
• Docs: Quickstart, Architecture, Payments (going live in India and abroad), Limits and Pricing, LLM Gateway, Deploy, Security.

The stack

FastAPI, async SQLAlchemy 2, Alembic, Postgres (SQLite for dev), optional Redis, official Anthropic and OpenAI SDKs. Next.js 15 (App Router), React 19, TypeScript, Tailwind CSS. Docker and Caddy.

How it works

1. Run it locally with the mock model and no keys (Quickstart).
2. Sign up, chat, hit the free cap, upgrade through the payment simulator, which sends signed webhook events through the real handler.
3. Add your Anthropic key, your Stripe and/or Razorpay keys, and follow PAYMENTS.md to go live.
4. Deploy to a VPS or a platform with migrations running before each swap.
5. Build your product on top: your prompts, your UI, your features, with your agent following AGENTS.md.

Who it's for

• Developers and founders building an AI assistant, tutor, copilot or chat SaaS who want to charge for it.
• Teams selling in India and abroad who need Razorpay and Stripe side by side.
• Builders who use AI agents and want them to start from a tested foundation.

Who it's not for

• No-code users. This is source code you run and own.
• Projects that need RAG, tool-using agents or a vector database out of the box (the model layer is ready for them; they are not included).
• Products that need team workspaces, seats or per-organization billing.
• Anyone who needs email verification, password-reset emails or SSO beyond Google included.

Proof

• 77 backend tests pass (pytest, 93 s), including parallel-request caps, duplicate webhook deliveries, refresh-token reuse and mid-stream failures.
• A 20-check end-to-end smoke run passes (backend/scripts/smoke_e2e.py, 22 s): signup, streaming chat, hitting the free cap, signed Razorpay webhooks, a repeated delivery changing nothing, an upgrade raising the limit, and the admin API refusing non-admins.
• Cold quickstart from a fresh copy: first streamed reply about 11 minutes after unzipping on a low-memory Windows laptop, almost all of it dependency installs.
• Real screenshots of every page are in screenshots/.

The patterns come from a live AI tutor that runs these mechanics in production: streaming runs, model fallbacks, daily fair-use caps, free follow-ups and re-explains, top-ups, Razorpay subscriptions with trials and webhooks, and memory of a user's mistakes.

FAQ

Do I need API keys to try it? No. The mock model and the payment simulator run everything locally. Add keys when you're ready.

Which models does it support? Anthropic (default) and OpenAI, plus any OpenAI-compatible endpoint through OPENAI_BASE_URL. Adding a provider is one class; LLM-GATEWAY.md shows how.

Can I use only Stripe, or only Razorpay? Yes. Leave the other provider's keys empty; the pricing page uses what is configured.

I'm an Indian business. Does Stripe work for me? Stripe's onboarding for India-registered businesses has been restricted (invite-only at the time of writing). Razorpay covers UPI, UPI Autopay and Indian cards; PAYMENTS.md has the go-live checklist.

Is it production-ready? It is production-shaped: Postgres, migrations, HTTPS deploy, signed and idempotent webhooks, a boot check that refuses unsafe production settings, and tests. You still own your keys, backups, monitoring, legal pages and email flows (see SECURITY.md).

Can my AI agent work on it? Yes. AGENTS.md, CLAUDE.md, the Cursor rule and the Claude Code skill tell it where things are, what must not break and how to verify changes.

What can I build with it? Any chat-shaped AI product: assistants, tutors, copilots, support bots. Change the system prompt, the plans and the UI.

What does the license allow? Unlimited commercial projects for you or your clients. You may not resell or share the starter itself. See LICENSE.md.

Do I get updates? Yes, for this version line. See CHANGELOG.md.