Vercel AI SDK: TypeScript Toolkit for AI Apps
AI SDK gives TypeScript projects one function call — generateText() — that reaches OpenAI, Anthropic, or Google models through Vercel's AI Gateway instead of three separate provider SDKs. Reach for it if you're already on Next.js, React, Svelte, or Vue and want to swap models by changing a string; skip it if your backend runs on Python or another non-JS stack, since every example in the README assumes TypeScript.
What is the Vercel AI SDK?
AI SDK is Vercel's TypeScript toolkit for connecting large language models to your code without locking into one vendor's SDK. You call generateText() with a model string like 'openai/gpt-5.4' or 'anthropic/claude-opus-4.6' and the library handles the provider-specific plumbing. The core package is published as ai on npm, with framework hook packages like @ai-sdk/react for chat interfaces.
Core Capabilities
- ✓Unified provider API — generateText() takes a model string like 'anthropic/claude-opus-4.6' or 'openai/gpt-5.4' and routes through the same function signature regardless of vendor.
- ✓Vercel AI Gateway by default — model calls resolve through Vercel's gateway unless you install a provider package like @ai-sdk/openai or @ai-sdk/anthropic and call the provider directly.
- ✓Structured output via Output.object() — pass a Zod schema and get back a typed object instead of parsing free-form text.
- ✓ToolLoopAgent class for building agents — wire tools like a shell executor or an image generator into a loop the SDK manages.
- ✓AI SDK UI hooks — useChat() and related hooks from @ai-sdk/react are framework-agnostic and work in Next.js, React, Svelte, and Vue.
- ✓createAgentUIStreamResponse() — streams an agent's output straight into a Next.js API route response.
Building AI-Powered Applications
- •Chat interfaces built with useChat() and a Next.js API route, matching the README's own image-generation-agent example.
- •Agents that call real tools — the README's sandboxAgent example gives a ToolLoopAgent shell access through Vercel Sandbox.
- •Structured-data extraction, like turning a prompt into a typed recipe object with ingredients and steps using generateText() plus Output.object().
- •Generative UI, where a tool call's result — such as a generated image — renders as a React component inside the chat stream.
Installation Steps
You need Node.js 22+ and npm, or another package manager; the README doesn't state a minimum npm version. Run npm install ai for the core package. If you'd rather call a provider directly than go through the Vercel AI Gateway, add its package too, e.g. npm install @ai-sdk/openai @ai-sdk/anthropic @ai-sdk/google. For UI hooks, install the framework package separately, e.g. npm install @ai-sdk/react. The README also suggests running npx skills add vercel/ai if you use a coding agent like Claude Code or Cursor.
Key Usage Patterns
Call generateText({ model: 'openai/gpt-5.4', prompt }) for a one-shot text response — the model string routes through the Vercel AI Gateway unless you've installed and imported a direct provider package. Swap in Output.object({ schema }) alongside a Zod schema to get a typed object back instead of prose, as the README's recipe example shows. For an agent, instantiate ToolLoopAgent with a model and a tools map, then hand it to createAgentUIStreamResponse() inside a Next.js route handler so useChat() in a client component can stream the conversation.
Strengths
- ✓One function signature (generateText) works across OpenAI, Anthropic, Google, and other providers — swapping models is a string change, not a rewrite.
- ✓AI SDK UI hooks like useChat() are framework-agnostic, so the same approach carries from Next.js to React, Svelte, or Vue.
- ✓Structured output via Output.object() with a Zod schema means TypeScript catches shape mismatches at compile time, not at runtime.
- ✓The npx skills add vercel/ai installer wires an AI SDK skill straight into coding agents like Claude Code or Cursor.
Considerations for Adoption
- △The README doesn't list a license, so check the repo directly before adopting this in a commercial codebase.
- △Every example in the README is TypeScript/JavaScript — there's no stated support for other languages, so this doesn't help a Python or Go backend.
- △Defaulting to the Vercel AI Gateway means your model calls route through Vercel's infrastructure unless you deliberately install and wire up a direct provider package.
- △The README's agent examples (ToolLoopAgent, Vercel Sandbox) assume you're comfortable building a tool-calling loop yourself — there's no low-code agent builder here.
Comparing AI Development Tools
Frequently Asked Questions
AI SDK's UI hooks work with Next.js, React, Svelte, and Vue, and the toolkit's broader docs also list Angular alongside Node.js as a supported runtime, per the GitHub README.
AI SDK connects to OpenAI, Anthropic, and Google out of the box through its unified provider API, plus additional providers listed in Vercel's provider docs; the README names these three explicitly.
The Vercel AI Gateway is the default routing layer AI SDK uses to reach model providers — pass a model string such as 'anthropic/claude-opus-4.6' and the gateway handles the rest, without installing a separate provider package.
AI SDK supports structured data generation through its Output.object() function, which pairs generateText() with a Zod schema so the response comes back as a typed object — the README's example generates a recipe with a name, an ingredients list, and steps.
AI SDK provides a ToolLoopAgent class that wires tools — like a shell command executor or an image generator — into a managed loop, and createAgentUIStreamResponse() streams that agent's output straight into a Next.js API route.
AI SDK templates are starter projects, per the README, that bundle AI SDK integrations for specific use cases, providers, and frameworks so you have a working setup to build from instead of starting blank.
The problem it solves
Every LLM provider ships its own SDK with its own request shape, streaming format, and tool-calling syntax, so a Next.js app that wants to let users pick between OpenAI and Anthropic models ends up maintaining two integration code paths and rewriting its chat UI's streaming logic for each one. AI SDK collapses that into one generateText()/useChat() surface so switching providers is a model-string change instead of a rewrite.
Who should try it — and who should skip
Reach for AI SDK if you're building a TypeScript app — Next.js, React, Svelte, or Vue — and want to swap between OpenAI, Anthropic, and Google models without rewriting your integration code each time. Skip it if your backend isn't TypeScript/JavaScript, since every README example assumes that stack, or if you want a visual, low-code agent builder instead of writing a ToolLoopAgent by hand.