TanStack AI

What is TanStack AI

TanStack AI is a type-safe, provider-agnostic, framework-agnostic TypeScript SDK for streaming chat, tool calling, structured output, media generation, MCP, and agents. The server talks to a model through an adapter, the client reads the stream through a hook for its framework, and the two halves speak the open AG-UI protocol. There is no hosted platform attached. It is built by the TanStack team behind TanStack Query and TanStack Router.

• Version: 1.0.0, published October 22, 2026 (announcement)
• Stability: 76 packages stable at 1.0; the agent harness is experimental at 0.x
• License: MIT
• Providers: 23 first-party adapters, plus openaiCompatible for any OpenAI-style API
• UI frameworks: React, Vue, Solid, Svelte, Preact, Angular, Octane, Remix, and a headless client for vanilla JS
• Community: 3 maintainers, 143 code contributors; AI Project of the Year at the 2026 JavaScript Open Source Awards
• Triage status: stable upstream, not yet used in an Aliz project

note

We covered the beta in June and the 1.0 release in October. This page is the evaluation summary; the posts have the detail.

Why it's interesting

• Stable core with a clear boundary. Every API deprecated in 0.x is gone, and 1.0 added no new concepts. The experimental parts are explicitly versioned 0.x.
• Protocol-first. AG-UI on the wire in both directions, so any AG-UI agent or client (LangGraph, CrewAI, Pydantic AI, a Go service) can sit on the other side of useChat.
• Per-model type safety. modelOptions, content types, and provider tools are typed from the model you pick. An option the model does not support fails at compile time.
• Isomorphic tools. toolDefinition() once, then .server() or .client(). Approval flows (needsApproval) arrive in the client as typed interrupts.
• Owns the hard parts. Persistence, resumable streams, compaction, memory, MCP client and server, sandboxes, Code Mode, and computer use are first-party packages, loaded only when imported.
• Framework-agnostic and platform-free. Eight UI frameworks plus a headless client; runs on Node, edge runtimes, TanStack Start, Next.js, React Router, Remix 3, Express, and React Native.
• Observable and testable. In-app devtools for server and client, OpenTelemetry middleware, and fakeText() for tests without a network or an API key.

How it fits together

Package	Role
@tanstack/ai	Core: chat(), decide(), streaming, the tool system, agent loop strategies, structured output, adapter types
@tanstack/ai-client	Headless chat client: message state, transports (SSE, HTTP stream, JSON, WebSocket, XHR), tool execution, approvals
@tanstack/ai-react and the -vue, -solid, -svelte, -preact, -angular, -octane, -remix packages	useChat and the generation hooks for your framework; createChatHook UI factories under the /ui subpath
@tanstack/ai-openai and 22 more adapters	One package per provider, tree-shakable per activity (openaiText, openaiImage, ...)
@tanstack/ai-persistence, @tanstack/ai-mcp, @tanstack/ai-sandbox-*, @tanstack/ai-compaction, @tanstack/ai-memory	Optional capabilities, each its own package
@tanstack/ai-harness, @tanstack/ai-harness-cli	Long-running agent sessions (experimental, 0.x)

The server route and the client hook are the whole integration surface for a chat:

api/chat.ts

import { chat, chatParamsFromRequest, toServerSentEventsResponse } from "@tanstack/ai";
import { openaiText } from "@tanstack/ai-openai";

export async function POST(request: Request) {
  const { messages, threadId, runId } = await chatParamsFromRequest(request);
  const stream = chat({ adapter: openaiText("gpt-6.1-sol"), messages, threadId, runId });
  return toServerSentEventsResponse(stream);
}

Chat.tsx

import { useChat, fetchServerSentEvents } from "@tanstack/ai-react";

const { messages, sendMessage, isLoading, stop } = useChat({
  connection: fetchServerSentEvents("/api/chat"),
  threadId: chatId, // required: a stable id per conversation
});

Switching providers is still an import change: anthropicText from @tanstack/ai-anthropic in place of openaiText, with the rest of the code untouched.

Capability map

Capability	Where	Status
Streaming chat, tools, agent loop	chat(), useChat, toolDefinition()	Stable
Structured output	chat({ outputSchema }), read as final in useChat; Zod, Valibot, ArkType, or JSON Schema	Stable
Typed decisions	decide() with choice, score, and boolean questions	Stable
Subagents	defineAgent() and chat({ subagents }), rendered as a card part	Stable
Tool approval and interrupts	needsApproval, interrupts[].resolveInterrupt()	Stable
Persistence and resumable streams	withPersistence middleware, durability adapters; useChat reconnects by itself	Stable
MCP client and server	@tanstack/ai-mcp, createMCPServer from @tanstack/ai-mcp/server, WebMCP in both directions	Stable
Media	Image, video, audio, speech, transcription, realtime voice, live worlds	Stable
Memory and compaction	memoryMiddleware (in-memory, Redis, mem0, Honcho, Hindsight), @tanstack/ai-compaction	Stable
Code Mode and computer use	Isolate drivers (Node, QuickJS, Cloudflare, Daytona, E2B); local, Playwright, and sandbox computer drivers	Stable
Sandboxes and coding agents	11 sandbox providers; Claude Code, Codex, OpenCode, Grok Build, and any ACP agent	Stable
Agent harness	defineHarness, createHarnessHost, runCli, createHarnessHandler	Experimental (0.x)

When to use

• AI features where provider portability matters: switching models or vendors should be an import change, not a rewrite.
• Backends that are not JavaScript. Anything that emits AG-UI (Python, Go, LangGraph, CrewAI, Pydantic AI) can sit behind useChat.
• Apps that need persistence, reload-safe streaming, or approval flows without hand-built routes.
• Teams already in the TanStack ecosystem (Query, Router, Start, Virtual).
• Projects where compile-time checks on model options and tool schemas are worth the typing discipline.

When NOT to use

• Apps that already speak the Vercel UI Message Stream (@ai-sdk/react, AI Elements, or a backend you do not control). TanStack AI does not read or write that protocol, so moving means rewriting the chat layer.
• Anything that depends on the agent harness for a production deadline. It is 0.x, can change in a minor release, and has documented recovery limits.
• Simple single-provider chat where our default Vercel AI SDK plus AI Elements already covers the UI and the team knows it.
• Teams that cannot keep client and server in lockstep. A 1.0 client needs a 1.0 server, and every stable package upgrades together.

Comparison to Vercel AI SDK

	TanStack AI	Vercel AI SDK
Maturity	Stable 1.0 (October 2026); harness experimental	Stable 7.x (7.0 since June 2026)
Ecosystem coupling	None, bring your own backend	Works anywhere, optimized for Vercel/Next.js
Protocol	Native AG-UI, in both directions	Own UI Message Stream; AG-UI through an external bridge
Type safety	Per model	Per provider
Tools	One toolDefinition() with .server() and .client()	tool() on the server, client tools through onToolCall
Persistence and resume	Built in	Patterns you write, plus the resumable-stream package
Providers	23 adapters plus openaiCompatible	About 38 first-party provider packages
License	MIT	Apache 2.0

TanStack's own comparison says to pick TanStack AI for a new app and to stay on the Vercel AI SDK when the app already speaks its protocol. Our Vercel AI SDK page and the AI Assistant Chat cookbook still default to the Vercel SDK. The maturity argument against TanStack AI no longer applies; the familiarity argument for the Vercel SDK still does.

Upgrading from 0.x

• Run the upstream codemod. It renames body to forwardedProps and conversationId to threadId.
• threadId is required on every chat and generation hook and on chat(). The codemod cannot add it.
• addToolApprovalResponse() is gone. Resolve the tool-approval interrupt instead.
• ai-react-ui, ai-solid-ui, and ai-vue-ui have no 1.0 release. Use createChatHook from @tanstack/ai-react/ui.
• Custom servers end a run with a RUN_FINISHED event, not data: [DONE].
• Stored chats, runs, and files need the saved-data steps from the upgrade guide before deploy.

caution

Upgrade every stable @tanstack/ai* package together and pin exact versions. The harness packages stay on 0.x.

Details are in the upgrade guide and in the upgrade section of our 1.0 post.

Open questions before it leaves triage

• A real project. Nobody at Aliz has shipped on it yet. The first candidate is a chat feature on a React SPA on Cloud Run, where SSE through the load balancer and a persistence store on our usual backends need to be proven.
• Bundle cost. Adapters are tree-shakable per activity. Measure the client bundle of @tanstack/ai-react plus @tanstack/ai-client against @ai-sdk/react on the same UI.
• Persistence on our stack. memoryPersistence() is for development. Check what a Firestore or Postgres store takes to implement against the store reference and its conformance suite.
• Overlap with Genkit. The cookbook pairs the Vercel SDK with Firebase Genkit on the server. Decide whether TanStack AI replaces Genkit flows or sits in front of them.
• Harness maturity. Re-evaluate the agent harness when it reaches 1.0.

Resources

• Official docs
• GitHub
• 1.0 announcement
• Upgrade to 1.0
• TanStack AI vs Vercel AI SDK
• Agent Skills for coding assistants
• Our posts: TanStack AI Beta and TanStack AI 1.0