Web Dev Tools

AI / LLM Tooling

Building with LLMs from a TypeScript / web app.

TypeScript SDKs / orchestration

  • Vercel AI SDK — provider-agnostic streaming, tool calling, structured outputs, generative UI. The default for new TS LLM apps in 2026.
  • Mastra — TypeScript framework for agents, workflows, RAG, evals. From the Gatsby founders.
  • LangChain.js — large, sometimes opinionated; mature retrieval/agent abstractions.
  • LlamaIndex.ts — strong RAG / index abstractions.
  • @anthropic-ai/sdk, openai, @google/generative-ai — provider-direct SDKs.
  • Ollama + ollama-js — local models.

Agent / orchestration frameworks

  • Inngest Agents — durable agent workflows on top of Inngest.
  • Trigger.dev with AI tools — durable LLM workflows.
  • CrewAI, AutoGen — multi-agent orchestration (Python-first; TS bindings emerging).

Vector / embeddings

  • pgvector + Postgres — for most apps, this is enough.
  • pgvectorscale — Timescale's vector index for huge corpora.
  • Pinecone, Qdrant, Weaviate, Milvus, Chroma — dedicated vector DBs.
  • Cloudflare Vectorize — Workers-native vector DB.
  • Turbopuffer, LanceDB — newer high-performance options.

LLM gateways / observability

  • Vercel AI Gateway, Cloudflare AI Gateway, Helicone, LangSmith, Langfuse, Braintrust, Arize Phoenix — pick at least one for production: caching, retries, traces, costs, evals.

Embeddings + chunking helpers

  • @upstash/ragify, LangChain text splitters, unstructured, Chonkie — chunking and ingestion.

Tool / agent UIs

  • assistant-ui — React components for chat UIs.
  • CopilotKit — embed AI copilots into existing apps.
  • Vercel AI SDK UIuseChat and friends; the default React chat hooks.

Local / on-device

  • Transformers.js — Hugging Face Transformers in the browser via WebGPU.
  • WebLLM — run LLMs in the browser via WebGPU.
  • MLX — Apple-silicon-native (TS bindings emerging).

Pick this if…

  • Default new project: Vercel AI SDK + your provider's API.
  • Agentic / multi-step durability: Mastra or Inngest Agents.
  • Heavy RAG focus: LlamaIndex.ts + pgvector (or Pinecone/Qdrant).
  • Need observability and caching: add an AI Gateway from day one.

On this page