AI & DevelopmentJavaScriptDeveloper Tools

TanStack AI 1.0: Stable APIs, 23 Providers, No Lock-In

TanStack AI 1.0 stable release - interconnected provider nodes and agent framework visualization
TanStack AI 1.0: Stable APIs, 23 Providers, No Lock-In

TanStack AI hit 1.0 on October 23. After going alpha in December 2025, beta in June, and release candidate in August, the framework that bet its identity on “no lock-in” AI tooling is now production-stable. The stable core ships alongside an experimental agent harness — 23 provider adapters, 8 UI frameworks, and a subagent loop that runs at version 0.x on purpose. ByteIota has covered TanStack AI through every milestone. This is the one that matters.

The Breaking Change You Need to Know First

If you’re upgrading from 0.x, the codemod handles most of it:

npx jscodeshift --parser=tsx -t https://raw.githubusercontent.com/TanStack/ai/main/codemods/ag-ui-compliance/transform.ts "src/**/*.{ts,tsx}"

That handles renames: conversationId becomes threadId, body becomes forwardedProps. What it cannot do is add threadId to every call — because it doesn’t know your IDs.

This is the real upgrade work. In 1.0, threadId is required on useChat, chat(), summarize(), and every generation hook. Neither client nor server auto-generates it for you anymore. You supply it from route params or your database:

const stream = chat({
  adapter: openaiText('gpt-6.1-sol'),
  messages: params.messages,
  threadId: params.threadId, // required — get this from your routes or DB
})

It’s a good architectural choice. Persistent conversation identity is table stakes for anything agent-shaped. The previous behavior — silently generating IDs — was always going to cause problems at scale. The migration is annoying; the result is correct.

What’s Actually New in 1.0

The headline numbers: the library grew from 4 to 23 named provider adapters, and from 3 to 8 UI framework bindings. New providers include Cloudflare Workers AI, Azure OpenAI, and Ollama for local inference. Frameworks now cover Vue, Svelte, Preact, Angular, Octane, and Remix alongside React and Solid. 143 contributors built this.

The more interesting additions are quieter. Computer use — where the model interacts with a web page via screenshots — ships with identical code across OpenAI and Anthropic providers. Automatic prompt caching is on by default, including for mid-conversation system prompt changes. Both of these matter for agent workloads more than they matter for chat boxes.

The Experimental Harness: Honest About What It Is

The agent harness (@tanstack/ai-harness) ships at version 0.x alongside the stable 1.0 core. That’s not a mistake. It means: “the chat primitives are frozen; the agent loop is not.”

What the harness brings: durable sessions that survive restarts via transaction logs, typed subagents with budgets (hand a turn to a named researcher or SEO agent, typed input enforced), approval interrupts that survive deployment, and a createHarnessMcpServer that exposes your harness as an MCP server — meaning Claude Code and Cursor can connect to it directly.

The subagent composition is worth understanding. A blog-writing harness example from the official 1.0 announcement runs three agents — research, draft, SEO — where an orchestrator picks which runs and can kick research and SEO off in parallel. This is a real multi-agent loop, not a marketing slide.

Shipping it at 0.x is the right call. It signals that the APIs may still change and that you shouldn’t build production infrastructure on top of them today. Teams burned by premature “stable” labels will appreciate the honesty.

Is It Time to Switch From Vercel AI SDK?

TanStack AI has 553K weekly npm downloads. Vercel AI SDK has 18.5M. It’s not a competition yet. But the technical case for TanStack is now concrete.

The bundle is 44% smaller (112 KB vs 199 KB gzipped). Under 100 concurrent streams, TanStack runs 29% faster (804ms vs 1,137ms) with 22% less memory. Per-model TypeScript typing catches errors at compile time that Vercel’s looser provider-level types let through. And the isomorphic tool definition pattern — write once, deploy with .server() or .client() — genuinely cuts tool code in half compared to Vercel’s required duplication. The independent benchmark from Upstash has the numbers.

The case against switching right now: Vercel AI SDK has pre-built agents available today; TanStack’s harness is still experimental. If you’re building something agent-heavy and need stability now, wait. If you’re building a chat feature, a streaming UI, or a multi-provider integration, 1.0 is ready. TanStack’s own comparison doesn’t pretend to win on every axis — worth reading before you decide.

Bottom Line

TanStack AI 1.0 is the stable foundation the project needed before it could credibly claim the “no lock-in” story. If you’re on 0.x, run the codemod, add your threadIds, and ship. If you haven’t looked at TanStack AI before, the 1.0 release PR on GitHub shows what changed in one place.

The agent harness is the part to watch. When it hits 1.0, the “no lock-in” pitch extends to the full agent loop. That’s the release that changes the comparison.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *