The Default AI / LLM Apps Stack
Here's the reframe that saves beginners weeks: an AI app is mostly a normal web app that calls a model. It's the web stack plus one API call to an LLM. Don't reach for exotic infrastructure — build the web app, add the model call, ship.
Read the Universal Toolbelt and web stack first — this page is just the AI-specific layer on top. And keep the rule above the stack in mind: buy or glue, don't build — call a hosted model instead of training your own, and rent the rest of the plumbing too.
| Slot | Pick | Why this one | Without it you'd... | Link |
|---|---|---|---|---|
| Model | Claude (Anthropic) — latest Opus/Sonnet | The latest Claude models (Opus 4.8, Sonnet 4.6) are top-tier at coding and reasoning, with excellent docs the AI knows well; start with Sonnet for cost, move to Opus for hard tasks. | Train and host your own model — months of work and GPU bills for something worse. | Anthropic API docs |
| App SDK | AI SDK (+ Anthropic SDK) | TypeScript-native streaming and chat UI helpers that turn a model call into a few lines, in the same Next.js app you already have — and it's vendor-neutral, so it runs on Cloudflare. | Hand-wire streaming responses and message state, and get the edge cases wrong. | AI SDK docs |
| Vector store | Supabase pgvector | Store embeddings for RAG in the same Postgres you already use — no second database vendor to learn. | Run a separate vector database service and sync data between two systems. | pgvector docs |
| Hosting | Cloudflare | Deploy the whole AI app — UI, API routes, and model calls — to a global edge network from one GitHub push. | Stitch together separate hosts for the front end and the model-calling back end. | Cloudflare Workers docs |
Don't bother yet
The AI space is loud and full of tooling that beginners over-adopt. Get one model call working first, then add complexity only when a real problem demands it.
- Heavy orchestration frameworks (LangChain, etc.) — abstractions you don't need yet; a direct SDK call is simpler and easier to debug. Add one only when your chains genuinely outgrow plain code.
- Fine-tuning — expensive and slow; good prompting plus RAG solves the vast majority of cases. Fine-tune only after you've proven prompting isn't enough.
- Self-hosted models — running your own model means GPUs and ops work; a hosted API is cheaper and better until you have serious scale or privacy requirements.
- A separate vector DB vendor — pgvector in your existing Supabase covers early RAG; don't add a Pinecone/Weaviate bill before you've hit its limits.
- Multi-agent setups — orchestrating many agents before a single model call works reliably is running before you can walk. Make one call solid first.