AI SDK Proxy

This project demonstrates how to use pydantic-ai's VercelAIAdapter to build a Python backend that is fully compatible with the Vercel AI SDK streaming protocol — letting a React frontend built with @ai-sdk/react talk to a FastAPI server backed by pydantic-ai agents.

What it demonstrates

VercelAIAdapter bridges two ecosystems: The Vercel AI SDK (TypeScript) and pydantic-ai (Python) use the same streaming wire format. VercelAIAdapter.dispatch_request handles translating an incoming useChat request into a pydantic-ai agent run and streaming the response back in the format the SDK expects.

Multi-provider switching at runtime: A single backend exposes two pydantic-ai agents — one backed by Anthropic (Claude Haiku) and one by OpenAI (GPT-4o mini). The frontend passes a ?provider= query parameter and the backend routes to the correct agent. No frontend changes are needed to support a new provider; you just add an agent.

Streaming markdown in React: Assistant responses are streamed token-by-token and rendered incrementally using streamdown, which handles markdown formatting as the stream arrives.

Architecture

Browser (React + @ai-sdk/react)
  │  useChat({ api: '/api/chat?provider=anthropic' })
  │
  ▼
Vite dev server  ──proxy /api──▶  FastAPI (port 8000)
                                      │
                                      │  VercelAIAdapter.dispatch_request()
                                      │
                                      ▼
                               pydantic-ai Agent
                               (anthropic | openai)

The Vite dev server proxies /api/* to http://localhost:8000, so the frontend never needs to know the backend's address.

Stack

Layer	Technology
Frontend	React 19, TypeScript, Vite, Tailwind CSS v4
AI SDK client	@ai-sdk/react (useChat), ai v6
Markdown streaming	streamdown
Backend	FastAPI, Python 3.12+
AI agent framework	pydantic-ai
LLM providers	Anthropic (claude-haiku-4-5), OpenAI (gpt-4o-mini)

Running locally

Backend

cd backend
cp .env.example .env          # add your API keys
uv sync
uv run fastapi dev main.py    # starts on http://localhost:8000

Frontend

cd frontend
pnpm install
pnpm dev                      # starts on http://localhost:5173

Open http://localhost:5173 and use the Anthropic / OpenAI toggle to switch providers mid-conversation.