Recommended stack · Good for startups

AI SaaS

A TypeScript-first application stack for streamed model features, retrieval, and an optional self-hosted inference path.

Responsibilities: 6
Open source: 6 choices
Self-hostable: 6 choices

What it is for

Start with a server-rendered product shell and provider-neutral AI SDK, add retrieval only when the product has a real data-grounding requirement, keep product state in PostgreSQL, and use a dedicated model server only when operating open models is justified.

Who should use it

Product teams building an AI feature or vertical AI application that need a pragmatic route from hosted APIs to owned model infrastructure.

From decision to first build

Start with this stack.

Choose the pieces you want, then copy the build prompt into your coding assistant. No account needed.

Optional capabilities start as “Not needed”. Hosted providers add external dependencies, including to the open-source stack.

Package manager: pnpm. All TypeScript setup, script, and CI instructions use pnpm.

What are you building?

Application framework

Next.js provides the server-rendered React surface and server boundary for streaming model interactions.

Open SourceRead the source ↗

Model integration

AI SDK offers typed, provider-neutral streaming, structured output, and tool-call primitives for TypeScript products.

Open SourceRead the source ↗

Retrieval layer

LlamaIndex adds ingestion and retrieval workflows when model responses must be grounded in product or customer data.

Open SourceRead the source ↗

Application data

PostgreSQL remains the system of record for users, entitlements, jobs, and auditable product state.

Open SourceRead the source ↗

Open-model serving

vLLM is the scale-oriented option when the team has evidence that operating open models is worth the GPU complexity.

Open SourceRead the source ↗

Error monitoring

Sentry captures failures around model calls, streaming boundaries, and the rest of the application release.

Generous Free TierRead the source ↗

Authentication

Application-owned authentication. Mount the Next.js App Router auth handler, configure a supported database adapter, and validate sessions on the server. Use nextCookies last in the plugin list if Server Actions set cookies.

Open-source core; you operate the database, email delivery, and authentication infrastructure.Read the source ↗

Styling

Add editable React component source using the official Next.js setup. Use Tailwind CSS and shared CSS-variable design tokens. Keep interactive components at client boundaries and preserve keyboard and screen-reader behavior when customizing.

Open-source components and styling; your team owns customization and upgrades.Read the source ↗

Identity verification (KYC) · Not needed

Identity verification (KYC)

Optional — add only when your product needs this capability.

Billing · Not needed

Billing

Optional — add only when your product needs this capability.

Transactional email · Not needed

Transactional email

Optional — add only when your product needs this capability.

File storage · Not needed

File storage

Optional — add only when your product needs this capability.

Background jobs · Not needed

Background jobs

Optional — add only when your product needs this capability.

Product analytics · Not needed

Product analytics

Optional — add only when your product needs this capability.

API contracts · Not needed

API contracts

Optional — add only when your product needs this capability.

Monorepo tooling · Not needed

Monorepo tooling

Optional — add only when your product needs this capability.

Agent integration · Not needed

Agent integration

Optional — add only when your product needs this capability.

Agent instructions · Not needed

Agent instructions

Optional — add only when your product needs this capability.

AI example · Not needed

AI example

Optional — add only when your product needs this capability.

CRUD example · Not needed

CRUD example

Optional — add only when your product needs this capability.

Data access / ORM · Not needed

Data access / ORM

Optional — add only when your product needs this capability.

Database hosting · Not needed

Database hosting

Optional — add only when your product needs this capability.

Documentation · Not needed

Documentation

Optional — add only when your product needs this capability.

Next.js includes React. The extra authentication and UI options are researched for Next.js; they disappear if you switch to another framework.

# Build with AI SaaS

Product brief: A TypeScript-first application stack for streamed model features, retrieval, and an optional self-hosted inference path.
Target user: Product teams building an AI feature or vertical AI application that need a pragmatic route from hosted APIs to owned model infrastructure.

## Selected technologies
- Language: TypeScript with strict checking.
- Package manager: pnpm (required). Use pnpm, not npm, npx, Yarn or Bun package-manager commands, for installation, scaffolding, scripts, examples, documentation and CI. Pin a compatible pnpm version in package.json packageManager, commit pnpm-lock.yaml, and use pnpm install --frozen-lockfile in CI. Use pnpm add, pnpm exec and the pinned version’s supported pnpm scaffolding command. Preserve existing work; do not delete or replace another lockfile without explicit migration approval.
- Package-manager reference: https://pnpm.io/installation
- UI library: React, through the Next.js App Router framework. React alone is not the full-stack framework.
- Application framework: Next.js
  Integration: Next.js provides the server-rendered React surface and server boundary for streaming model interactions.
  Cost context: Open Source
  Reference: https://howtobuild.dev/projects/nextjs
- Model integration: AI SDK
  Integration: AI SDK offers typed, provider-neutral streaming, structured output, and tool-call primitives for TypeScript products.
  Cost context: Open Source
  Reference: https://howtobuild.dev/projects/ai-sdk
- Retrieval layer: LlamaIndex
  Integration: LlamaIndex adds ingestion and retrieval workflows when model responses must be grounded in product or customer data.
  Cost context: Open Source
  Reference: https://howtobuild.dev/projects/llamaindex
- Application data: PostgreSQL
  Integration: PostgreSQL remains the system of record for users, entitlements, jobs, and auditable product state.
  Cost context: Open Source
  Reference: https://howtobuild.dev/projects/postgresql
- Open-model serving: vLLM
  Integration: vLLM is the scale-oriented option when the team has evidence that operating open models is worth the GPU complexity.
  Cost context: Open Source
  Reference: https://howtobuild.dev/projects/vllm
- Error monitoring: Sentry
  Integration: Sentry captures failures around model calls, streaming boundaries, and the rest of the application release.
  Cost context: Generous Free Tier
  Reference: https://howtobuild.dev/projects/sentry
- Authentication: Better Auth
  Integration: Application-owned authentication. Mount the Next.js App Router auth handler, configure a supported database adapter, and validate sessions on the server. Use nextCookies last in the plugin list if Server Actions set cookies.
  Cost context: Open-source core; you operate the database, email delivery, and authentication infrastructure.
  Reference: https://better-auth.com/docs/integrations/next
- Styling: shadcn/ui + Tailwind CSS
  Integration: Add editable React component source using the official Next.js setup. Use Tailwind CSS and shared CSS-variable design tokens. Keep interactive components at client boundaries and preserve keyboard and screen-reader behavior when customizing.
  Cost context: Open-source components and styling; your team owns customization and upgrades.
  Reference: https://ui.shadcn.com/docs/installation/next

## Implementation brief
1. Inspect the current repository and its instructions. Explain the architecture and verify the current official installation guides and package compatibility for the selected choices. Preserve existing work.
2. Implement one useful vertical slice with a responsive, accessible interface, loading/empty/error states, and a clear README. Use only the selected provider for each responsibility. Do not silently substitute another stack.
3. If authentication is selected, enforce authorization on the server. Implement sign-in, sign-out, session handling, and protected data access. Keep provider secrets server-side and document required environment variables.
4. Document data models and migration commands. Do not provision paid services, deploy, or modify an existing database without explicit authorization. Use placeholders in .env.example and never invent API keys.
5. Validate with the repository’s allowed checks and record their results. Document what remains dependent on account setup, credentials, or deployment.
6. Do not add optional providers that were not selected. For external integrations, keep secrets server-side, verify webhook signatures, and make event handling idempotent. Hosted add-ons change the base stack’s self-hosting and cost assumptions.

## Base stack tradeoffs (reassess when a choice is replaced)
- AI model cost and latency remain product constraints no framework removes.
- A retrieval framework is unnecessary complexity for features that do not need grounded private data.
- Self-hosted inference requires GPU operations expertise and should not be adopted by default.
- These selections customize the editorial stack. Recheck integration boundaries when replacing a component; hosted free tiers are not a promise of free production operation.

## Stack references
- https://howtobuild.dev/stacks/ai-saas
- Provider-neutral streaming and tool integration capabilities: https://ai-sdk.dev/docs
- High-throughput open-model serving characteristics: https://docs.vllm.ai/en/latest/

Updates as you choose your technologies.

Responsibility map

One choice for each job.

01
Application framework
Next.js
Next.js provides the server-rendered React surface and server boundary for streaming model interactions.
Open sourceOpen SourceSelf-hostable
02
Model integration
AI SDK
AI SDK offers typed, provider-neutral streaming, structured output, and tool-call primitives for TypeScript products.
Open sourceOpen SourceSelf-hostable
03
Retrieval layer
LlamaIndex
LlamaIndex adds ingestion and retrieval workflows when model responses must be grounded in product or customer data.
Open sourceOpen SourceSelf-hostable
Alternatives
PostgreSQL — Use PostgreSQL directly when modest retrieval needs do not justify a separate framework.
04
Application data
PostgreSQL
PostgreSQL remains the system of record for users, entitlements, jobs, and auditable product state.
Open sourceOpen SourceSelf-hostable
05
Open-model serving
vLLM
vLLM is the scale-oriented option when the team has evidence that operating open models is worth the GPU complexity.
Open sourceOpen SourceSelf-hostable
Alternatives
Vercel — Stay with hosted model APIs and managed deployment when GPU operations are not a product advantage.
06
Error monitoring
Sentry
Sentry captures failures around model calls, streaming boundaries, and the rest of the application release.
Open sourceGenerous Free TierSelf-hostable

Open-source profile

What you can own.

The application framework, AI libraries, database, and model server are open source. Model providers and the selected deployment and monitoring services may be proprietary.

Cost profile

Where cost appears.

Early prototypes can use free tiers, but model tokens, vector or relational storage, GPU inference, observability, and abuse controls become usage-sensitive costs.

Tradeoffs

What this stack asks you to accept.

AI model cost and latency remain product constraints no framework removes.
A retrieval framework is unnecessary complexity for features that do not need grounded private data.
Self-hosted inference requires GPU operations expertise and should not be adopted by default.

Editorial note

Stack choices are editorial recommendations, not an automatic popularity ranking. Pricing and capabilities can change.

Sources checked

Provider-neutral streaming and tool integration capabilitiesSep 4, 2026

High-throughput open-model serving characteristicsSep 4, 2026