TanStack AI 1.0: 76 Packages, 23 Providers, Zero Lock-In

Less than a year ago, TanStack AI was a single chat() function, four provider adapters, and one promise: no lock-in, ever. On October 23, 2026, that experiment became a 1.0 — TanStack's formal commitment that the APIs are stable and ready for production.

The numbers behind the release: 76 packages went stable together (per TanStack's own docs index), adapters grew from 4 to 23 — plus an openaiCompatible catch-all for any OpenAI-style API — across eight UI frameworks, with 143 contributors on a project the maintainers call "still built by a small team of three." The price is the whole pitch: MIT license, free, no paid tiers. Nothing to upsell.

Here's what shipped, what's stable versus experimental, what the free license really costs in practice, and how to start or upgrade — every technical claim below comes from the official TanStack AI 1.0 announcement, published October 23, 2026.

Quick answer

• What it is: an open-source, provider-agnostic TypeScript toolkit for AI features — streaming chat, tool calling, media generation, MCP, sandboxes, and agents — that runs on any UI framework and any model provider.
• What's new in 1.0: stable core APIs, 76 packages released together, an experimental agent harness, fully rewritten docs, and a one-codemod upgrade path from 0.x.
• Price: free. MIT license, no paid tiers, no commercial roadmap. You still pay your model providers — the toolkit just stops you from paying for a framework.
• Where to start: the official Quick Start promises a streaming chat in about five minutes from a single npm install.
• The catch: the new agent harness ships as experimental 0.x — its APIs can still change in a minor release.

What TanStack AI 1.0 actually is

TanStack AI is the AI arm of the TanStack project — the same people behind TanStack Query, Table, and Start. Where TanStack Start 1.0 (shipped October 13, and already the default framework for Netlify's Agent Runners, per Firerun) handles full-stack web apps, TanStack AI handles the AI layer: streaming chat, tool calls, media generation, MCP, sandboxes, and agents.

The design thesis hasn't changed since the December 2025 alpha: every layer is swappable — provider, framework, sandbox, wire protocol. TanStack backs the stability claim with an unusually honest testing story: end-to-end tests run on every pull request against deterministic model fixtures, no live API calls. Provider quirks are the silent killer of "universal" SDKs, and fixture testing is how 23 adapters stay honest. The payoff: one conversation can replay across models, so a chat started on a cheap provider can move to a frontier one mid-conversation.

The pricing fine print: free, with one asterisk

TanStack's own site lists its terms plainly: license MIT, hosting self-host anywhere, governance community-driven, paid tiers: none. The npm registry confirms @tanstack/ai at version 1.0.0 under an MIT license, and the RC announcement states there is no commercial roadmap, no product to upsell, no hidden agenda. Community pull requests are the business model.

Here's the asterisk, worth stating outright because "free" misleads: the toolkit charges you nothing, but your models still do. TanStack AI ships bring-your-own-key (defineByok, keyed adapters) and sign-in with ChatGPT, GitHub Copilot, OpenRouter, or a Claude OAuth token — so users of a local app don't have to paste an API key. Every one of those routes ends at a provider's billing page. The honest math: the framework costs $0 forever; your token bill depends on which provider you point the adapters at — and since swapping providers doesn't mean rewriting the app, you can finally shop them on price.

What's stable, what's experimental

The most important sentence in the 1.0 announcement: "1.0 has no new concepts." This is a stability promise, not a feature drop.

Since the August RC, the toolkit added subagents in chat() and useChat (hand turns to a named child agent with typed input and limits), decide() for typed decisions, portable agent skills, chat compaction, computer use (the model drives a web page from screenshots — same code on OpenAI and Anthropic), automatic prompt caching, and fakeText() for offline testing with no API key. MCP goes both directions: serve your tools as an MCP server, or let browser agents call your client tools via WebMCP.

The experimental harness, explained honestly

chat() answers one request — fine for a chat box. A coding agent, a support bot, or a media studio needs more: one conversation that stays open, takes new messages while it works, runs other agents, and survives a deploy. So TanStack built the harness: define it once with the same options as chat(), open a session per conversation.

Sessions give you durable turn logs (a turn cut by a crash continues on the next host), waits that survive restarts, typed agents with budgets and timeouts, multi-user access rules, and background workers. The run route speaks standard AG-UI events over HTTP or WebSocket, and first-party plugins cover file tools, permissions, snapshots, todos, and usage tracking.

Two caveats, both from the announcement. The harness APIs are experimental and can change in a minor release — TanStack shipped it early on purpose. And read this footnote twice: the permissions plugin "isn't a sandbox — these tools run on the host with its permissions." If your agent touches the real filesystem, the sandbox decision is still yours.

Quickstart: a streaming chat in about five minutes

The official Quick Start promises a streaming chat in roughly five minutes from one install — core package verified on npm as @tanstack/ai 1.0.0 (MIT):

1. Install. npm install @tanstack/ai, plus your provider adapter and UI package. If OpenAI isn't your provider, swap the provider package; if React isn't your framework, swap the React package for one of the other seven.
2. Add your key — or don't. Bring your own key, or let users sign in with ChatGPT, GitHub Copilot, OpenRouter, or a Claude OAuth token. For local testing, fakeText() exercises the whole pipeline offline with no key.
3. Build the UI headlessly. The old @tanstack/ai-react-ui-style packages are gone; 1.0 ships typed, headless chat UI factories — createChatHook from @tanstack/ai-react/ui, with Solid and Vue equivalents and custom adapters for any framework.
4. Follow the rewritten docs. Every page opens with the problem and a setup you can copy and run in a few minutes — including agent skills so Claude Code or Cursor writes correct TanStack AI code for your version.

Upgrading from 0.x: one codemod, one caveat

If you kept up with the deprecation warnings, the announcement says you're most of the way there. The upgrade: run the codemod (renames body → forwardedProps, conversationId → threadId), update all stable @tanstack/ai* packages to 1.0 together (a 1.0 client needs a 1.0 server; harness packages stay on 0.x), remove the old @tanstack/ai-*-ui packages, run your type checker, and do the saved-data steps before deploying if you keep data from 0.x.

The caveat: threadId is now required on the browser and the server, and neither side generates one for you — the codemod can't add what it doesn't know. The type check also won't migrate stored chats, runs, or files, so follow the saved-data steps before deploying or old sessions won't load.

Why this release matters beyond TanStack

• Distribution is already here. Netlify made TanStack Start 1.0 the default for new Agent Runners projects on October 13 — agent-builder scaffolds now ship on this stack by default. Vercel is a Gold TanStack partner with integrations for Start, AI Gateway, and Sandbox.
• Agents are the whole industry's problem. This week has seen OpenAI's Decisions API go wide and Anthropic ship its first personal agent. An SDK that treats subagents, sandboxes, and computer use as first-class citizens — not a chat box with extras — matches where demand actually is.
• The open-weights current. The same open-source momentum is running through the model layer — see our coverage of Recursive's free 40B laptop coding model and DeepSeek's open-source inference engine. TanStack AI is the framework-layer version of that story.

Pros and cons

Pros:

• Genuinely provider-agnostic with compile-time type safety — provider/model mismatches fail at compile time, not in production.
• An unusually honest testing story: deterministic E2E fixtures on every PR, no live API calls needed.
• Free forever with no paid tiers and no upsell — stated outright in the RC announcement — plus fully rewritten docs with copy-runnable setups.

Cons:

• The agent harness — the most interesting new piece — is experimental and can change in a minor release.
• Three maintainers for 76 packages is an honest bus-factor concern, stated by the project itself.
• "Free" covers the framework only; provider API spend is untouched, and adapter parity is only as good as the fixture tests.

Alpha vs 1.0: ten months of progress

(Adapted from the official 1.0 announcement.)

FAQ

Is TanStack AI really free?

Yes, with one asterisk: the framework is MIT-licensed with no paid tiers and no upsell. But your model providers still bill you — the savings are in framework freedom, swapping providers on price without rewrites.

What does "stable" actually cover?

The core APIs — chat, tools, media, MCP, persistence, sandboxes, and adapters — are production-stable. The exception is the agent harness (@tanstack/ai-harness), experimental 0.x and allowed to change in a minor release.

Can I switch LLM providers without rewriting my app?

That's the design goal. Swap the adapter package (or use openaiCompatible for any OpenAI-style API) and keep the rest — chat history replays across models, so one conversation can move providers.

Which frameworks and providers are supported?

Eight UI frameworks: React, Vue, Solid, Svelte, Preact, Angular, Octane, Remix, plus vanilla JS. Twenty-three provider adapters, including OpenAI and Anthropic plus newer additions like Cloudflare Workers AI, Azure OpenAI, Lovable AI Gateway, World Labs Marble, TypeSafe, and Ollaya — with OpenRouter a partner since the beginning.

Should I upgrade from 0.x right now?

If you ship to production, yes — 1.0 is the version TanStack asks you to build on. The upgrade is one codemod plus a type check; just upgrade all stable @tanstack/ai* packages together and handle the new required threadId — neither side generates it for you.

Key takeaways

• TanStack AI 1.0, shipped October 23, 2026, is a stability release — "1.0 has no new concepts," and that's the point.
• The framework is MIT-licensed and free with no paid tiers; your provider API bills are untouched, but you can shop providers on price.
• 23 provider adapters plus an openaiCompatible catch-all and 8 UI frameworks mean provider switches without rewrites.
• The upgrade: one codemod — upgrade all packages together, and plan for the new required threadId.
• The experimental agent harness is the part to watch, not adopt: it can change in a minor release, so pin it if you touch it.

TanStack AI 1.0 doesn't ask you to believe in a roadmap — it asks you to build on APIs that finally come with a commitment attached. If you've been waiting for a sign to build AI features without picking a provider lock-in first, this was it.