★ AI agent · stack
Best Tech Stack for an AI agent MVP
My opinionated stack for an AI agent MVP, layer by layer, with why I'd pick each tool and what would change my mind. Adjust the answers to fit your team.
- Runs in your browser
- No signup
Your stack
Tech stack for AI agent
9 layers, picked for your answers. Boring where it should be, so your effort goes into the product.
- AI features: log every prompt and response, keep an eval set, and cap spend per user from day one.
- Frontend
Next.js (React + TypeScript)
Server rendering for SEO, file-based routing and the biggest React ecosystem. TypeScript end to end catches whole classes of bugs before users do.
- What would change my mind
- A purely internal tool with no SEO needs, where a plain Vite + React single-page app is simpler.
- Alternative
- Vite + React SPA, or Remix / React Router
- Backend / API
Next.js route handlers + server actions
No separate service to deploy: the API lives next to the UI and shares types with it. Plenty for an MVP and well beyond.
- What would change my mind
- Several clients (mobile, partners) hitting the API, long-running work, or a separate backend team.
- Alternative
- A standalone Node.js + Fastify service
- Database
PostgreSQL (managed: Neon or AWS RDS)
Relational, transactional and boring. JSON columns cover the semi-structured bits, and pgvector covers embeddings if you add AI.
- What would change my mind
- Truly schemaless, write-heavy event data at very high volume. Add a specialised store for that slice only.
- Alternative
- MySQL, or MongoDB for document-shaped data
- Auth
Clerk
Prebuilt sign-in UI, organisations, roles and MFA. Hours instead of weeks for a B2B-ready login.
- What would change my mind
- Per-user pricing at consumer scale, or a hard requirement to keep auth in-house.
- Alternative
- Auth.js (NextAuth)
- Hosting
Vercel
Git push to deploy, preview URLs for every branch and no servers to manage, which is exactly what an MVP team needs.
- What would change my mind
- Long-running processes, heavy compute or strict data-residency requirements.
- Alternative
- Netlify, or Railway for a container
- AI layer
Vercel AI SDK behind your own wrapper + pgvector
One thin internal interface over the model API, so you can switch or route models as prices change. Embeddings live in pgvector, inside the Postgres you already run.
- What would change my mind
- Tens of millions of vectors or heavy filtering, which is when a dedicated vector database earns its keep.
- Alternative
- A dedicated vector database (Qdrant, Pinecone)
- Background jobs
Inngest or Trigger.dev
Durable background functions with retries, schedules and a dashboard, written in the same TypeScript codebase.
- What would change my mind
- Very high throughput or strict data residency, which points to a queue in your own cloud.
- Alternative
- BullMQ + Redis
- Analytics
PostHog
Product analytics, funnels, session replay and feature flags in one tool, with a generous free tier.
- What would change my mind
- A marketing-only site, where a simple privacy-friendly counter like Plausible is enough.
- Alternative
- Plausible (simple), or Mixpanel
- Email
Resend
Reliable transactional email with a clean API and good deliverability defaults.
- What would change my mind
- Marketing campaigns and newsletters, which belong in a separate tool.
- Alternative
- Postmark
This is the stack I'd pick.
Want me to build on it? I ship MVPs on exactly this kind of stack.
What makes an AI agent different
An agent takes actions, so the architecture is about control: durable workflows that survive failures, human approval for risky steps, and a log of every tool call so you can explain what happened.
Risks to plan for
Unsupervised actions
Anything touching money, customers or deletion needs an approval step and an audit trail.
Fragile long runs
Multi-step runs fail halfway. Use durable background functions with retries rather than a single request.
Silent regressions
A model update can change behaviour overnight. Keep scenario tests and run them on every change.
The stack at a glance
With the preset answers for an AI agent (you can change them above), this is the stack layer by layer.
| Layer | Pick | Alternative |
|---|---|---|
| Frontend | Next.js (React + TypeScript) | Vite + React SPA, or Remix / React Router |
| Backend / API | Next.js route handlers + server actions | A standalone Node.js + Fastify service |
| Database | PostgreSQL (managed: Neon or AWS RDS) | MySQL, or MongoDB for document-shaped data |
| Auth | Clerk | Auth.js (NextAuth) |
| Hosting | Vercel | Netlify, or Railway for a container |
| AI layer | Vercel AI SDK behind your own wrapper + pgvector | A dedicated vector database (Qdrant, Pinecone) |
| Background jobs | Inngest or Trigger.dev | BullMQ + Redis |
| Analytics | PostHog | Plausible (simple), or Mixpanel |
| Resend | Postmark |
Want the price tag too? Run the same product through the AI agent cost estimate.
Questions founders actually ask
What's the most important layer for an AI agent?
Do I need an agent framework?
How do I keep an agent safe?
Why does the picker recommend boring technology?
Is this stack good enough to scale?
Stack by product type
- Stack for a B2B SaaS
- Stack for a marketplace
- Stack for an AI chatbot
- Stack for a mobile app
- Stack for a fintech app
- Stack for an e-commerce store
- Stack for a booking app
- Stack for a CRM
- Stack for a learning platform
- Stack for a social / community app
- Stack for an internal tool
Or start from a blank slate with the full tech stack picker.
More from the workbench
Fine print
Tools: on the house.
Engineer: by the hour.
The opinion was free. The implementation is where I come in.
I'm Himanshu, an AI and full-stack engineer. I build MVPs, AI agents and the growth plumbing around them for early-stage startups.