
TanStack AI 1.0 shipped October 23 with 76 packages stable, 23 provider adapters, and an experimental agent harness that routes coding work to Claude Code, Codex, or any ACP-compatible agent in an isolated sandbox. The library that started as a single chat() function with four providers has become what most developers have been building themselves — a provider-agnostic AI layer that doesn’t break when a model gets sunsetted.
The Provider Count Is the Real Story
TanStack AI went from four alpha providers to 23 stable adapters, plus an openaiCompatible escape hatch for anything implementing the OpenAI Chat Completions wire format. New additions include Cloudflare Workers AI, Cloudflare AI Gateway, Azure OpenAI, Lovable AI Gateway, World Labs Marble, and Ollaya. Five new sandbox providers — Blaxel, Upstash Box, E2B, boxd, and Railway — round out the ecosystem.
That number matters this week specifically. Google stopped selling new Gemini Code Assist subscriptions on October 9. Firebase Studio stops accepting new workspaces entirely and shuts down March 2027. Developers who hard-coded their AI integrations are scrambling. With TanStack AI, a provider sunset means changing one import.
The openaiCompatible adapter extends coverage further. Any API speaking the OpenAI Chat Completions wire format — self-hosted models, regional providers, internal proxies — works without a dedicated adapter package. You install only what you use; the rest doesn’t ship to your bundle.
The One Breaking Change You Must Handle Yourself
The 1.0 migration is designed to be smooth — deprecated 0.x APIs already had replacements in place before the cut. But there is one change the provided codemod cannot handle: threadId is now required everywhere.
In 0.x, TanStack AI auto-generated a thread ID if you didn’t provide one. In 1.0, you own it:
// 0.x — breaks silently in 1.0
useChat({ connection })
// 1.0 — threadId required
useChat({ connection, threadId: chatId })
This applies to useChat, chat(), summarize(), every generation hook, and every generation function. On the server, conversationId is now threadId. Pass a stable ID you control — a route ID, a database record ID, anything that persists across sessions. The codemod handles the other removals (body renamed to forwardedProps, addToolApprovalResponse replaced with an interrupts pattern), but it cannot infer your IDs.
Run the codemod, then audit every callsite that used useChat or chat() manually. The official migration guide walks through each change.
The Experimental Agent Harness
The headline new addition is @tanstack/ai-harness, an experimental package for running long-lived, stateful AI agents beyond a single request. It supports durable sessions that survive restarts, typed sub-agents running in parallel or series, budgets, timeouts, MCP, and background work.
The concrete use case: hand a coding task to Claude Code, Codex, OpenCode, or Grok Build — each running in its own sandbox, streaming output back to your app. Any Agent Client Protocol (ACP) agent works through the generic harness adapter. Supported sandboxes include Blaxel, Upstash Box, E2B, boxd, and Railway.
It is explicitly experimental. The ACP protocol is still evolving, and multi-user session management patterns are being worked out. Do not build production workflows on it today. But this is the direction: instead of calling a stateless API, your app orchestrates agents with state, budgets, and isolation. Watch this package.
TanStack AI or Vercel AI SDK?
Both are stable now. Vercel AI SDK reached 7.0 stable in June 2026; TanStack AI reached 1.0 this week. The choice has crystallized:
- Pick TanStack AI if you need Solid or Preact support, want per-model type safety, deploy to Cloudflare Workers, or want a server bundle that’s 44% smaller than the AI SDK baseline. Read the official comparison for the full breakdown.
- Pick Vercel AI SDK if you are deep in Next.js and want a stable agent harness today, not experimental.
If you are not locked into the Next.js/Vercel deployment stack, TanStack AI 1.0 is now the defensible default. The 23-provider ecosystem, framework agnosticism, and smaller bundle make it the choice that ages better as the provider landscape keeps shifting.
Getting Started
npm install @tanstack/ai @tanstack/ai-react @tanstack/ai-openai
Swap the adapter package for your provider — @tanstack/ai-anthropic, @tanstack/ai-cloudflare, @tanstack/ai-azure-openai, and so on. Full details in the TanStack AI documentation.
TanStack AI won AI Project of the Year at the 2026 Open Source Awards. With 735,616 weekly downloads and 143 contributors, it is past the “interesting experiment” phase. The AI feature your app needs six months from now probably runs on a provider that doesn’t exist in your current integration. Build on infrastructure that handles that gracefully. Read the full 1.0 release post for everything that shipped.

