The 2026 SaaS Stack We Actually Ship On

We ship MVPs in 28 days. Not “agile sprints” or “rapid prototyping”-actual, live, paying-customer-ready software. That timeline forces choices. You can’t afford tech debt, religious debates about frameworks, or tools that look good in a demo but cost you two weeks of integration hell.

This is what we actually use at Amora in 2026, why we chose it, and what we’ve deliberately avoided.

Frontend: Next.js and Vercel, with Shadcn/ui

React isn’t new. But Next.js 15 with App Router is the fastest way to ship a full-stack application without context-switching between your API layer and your UI. Server components, incremental static regeneration, and built-in API routes mean you’re not maintaining two separate codebases.

We build on Vercel’s platform. Yes, it costs more than self-hosted. No, we don’t regret it. Deploy on git push. Edge functions run in 200+ regions worldwide. Analytics and monitoring come bundled. For a 28-day build, that saves us a week of infrastructure bikeshedding.

Shadcn/ui is our component library. It’s not a package you import; it’s code you copy into your repo and own. That matters when you need to modify a button behaviour at 4 PM on day 27. Tailwind CSS underneath. No CSS-in-JS runtime overhead.

Alternatives we’ve tested and abandoned: Vue (smaller ecosystem), Astro (overkill for interactive dashboards), custom Svelte (great, but fewer developers familiar with it when you need to scale the team).

Backend: PostgreSQL, Node.js, and Prisma

PostgreSQL is the only database we reach for. It’s been stable since 2005. It handles JSON natively, scales vertically to handle most SaaS workflows, and you’re never locked into a cloud vendor. We run it on Supabase for early products (managed, AUD 200-800/month depending on usage), then move to AWS RDS or Railway for control.

Node.js on the backend because your frontend engineers can write backend code, and you cut hiring complexity. TypeScript everywhere-not optional, mandatory. The type safety catches bugs before you deploy, and it makes refactoring at speed actually possible.

Prisma as the ORM. It generates type-safe database clients from your schema, migrations are declarative and reversible, and the developer experience is noticeably better than raw SQL queries or competing ORMs. A fintech we worked with switched from Sequelize mid-project; the team got 40% faster.

We’ve ditched: ORMs with excessive magic (Waterline), non-relational databases for transactional products (sorry MongoDB enthusiasts-ACID guarantees matter), and microservices at MVP stage (deploy a monolith, split later if you actually need to).

AI and Inference: Claude API and Replicate

We call Claude’s API directly via Anthropic’s SDK. Anthropic’s pricing is clear (AUD 0.003 per 1k input tokens, AUD 0.015 per 1k output tokens, rough ballpark), the models are strong, and they’re responsive when you report issues.

For vision (processing images, PDFs, screenshots), Claude 3.5 Sonnet is good enough for most products. If you need faster inference or specific model control, Replicate lets you run open-source models (Llama, Flux, Whisper) on GPUs without managing infrastructure.

What we avoid: self-hosting LLMs unless the founder has strong reasons (data privacy, cost at massive scale). The operational overhead-GPU provisioning, model fine-tuning, version management-kills your 28-day timeline. Use the API until you have hundreds of thousands of requests per month and the budget to justify it.

Most founders assume they need fine-tuning. They don’t. Prompt engineering, RAG (retrieval-augmented generation), and system prompts get you 80% of the way. Fine-tuning is a later-stage optimisation.

Infrastructure and DevOps: Railway or AWS, minimal tooling

For simplicity: Railway. Deploy from git, environment variables are encrypted, databases are managed, and the pricing is transparent. You pay for compute time and storage. No surprise AWS bills at 2 AM.

For control: AWS. We use EC2 (or ECS for containerised apps), RDS for databases, S3 for file storage, and CloudFront for CDN. More complex, but cheaper at scale and you own the keys.

We don’t use Kubernetes. Seriously. It’s incredible at 1,000+ engineers and millions of daily users. For an MVP, it’s overhead masquerading as architecture. Docker containers on Railway or simple ECS suffice.

Monitoring and logging:

• Sentry for error tracking. Catches bugs in production you’d never spot otherwise.
• Datadog or Axiom for logs. Axiom is cheaper (AUD 50-150/month for early-stage), Datadog is the enterprise standard.
• Uptime monitoring: Postman or Checkly. Five-minute health checks prevent surprises.

We avoided Splunk (enterprise cost), ELK stack (maintenance burden), and vendor-specific monitoring (ties you in too early).

The Non-Tech Layer: Stripe, SendGrid, and Auth0

Stripe for payments. Full stop. Their API is clean, their fraud detection actually works, they handle compliance, and their dashboard is useful. Even for non-financial SaaS, they’re a better alternative to custom payment processing.

Email: SendGrid or AWS SES. SendGrid (AUD 30-100/month) if you need templates and deliverability support. SES if you’re sending high volume and want the cheapest per-email cost.

Authentication: Supabase Auth (free, built in if you’re using Supabase for the database) or Auth0 (AUD 0-150/month depending on user count). Don’t build authentication yourself. Full stop. Not for 28 days, not ever. The compliance, the security, the edge cases-it will eat your timeline.

Analytics: Posthog or Mixpanel. Both offer event tracking without the privacy concerns of Google Analytics. Posthog is cheaper and open-source.

What Changed Since 2024

TypeScript adoption hit critical mass. We now require it everywhere; finding developers is no longer the blocker it was two years ago.

AI inference is genuinely cheap. Six months ago, we hedged bets across OpenAI, Anthropic, and open-source models. Now? Claude is fast and stable enough that we don’t need to branch logic.

Vercel’s edge functions and middleware actually work. Request-level processing at the edge is useful now, not a future promise.

How This Matters for Your Build

Every tool here was chosen because it gets out of your way. You’re not spending time learning Terraform syntax, configuring databases by hand, or debugging deployment pipelines. You’re building features.

If you’re building a SaaS product, an AI agent, or a web platform and you’re not sure whether to custom-build or use existing services, the answer is almost always “use existing services.” The money you save on infrastructure complexity is money you can spend on design, on your core product, or on user acquisition.

If you want to talk through the specifics for your project-whether this stack makes sense, where you might need to deviate, or how to plan a 28-day build-talk to Amora about your build. We do this for a living.

The companies winning in 2026 aren’t the ones with the most sophisticated infrastructure. They’re the ones shipping fast, learning from users, and iterating. This stack lets you do exactly that.