TanStack AI 1.0: Stable Core, 23 Providers, Experimental Harness

TanStack AI hit 1.0. The open-source TypeScript toolkit for chat UIs, tool calling and agents went stable on October 23 with 23 provider adapters, support for nine UI frameworks, and a pitch we like: the model provider is your choice, not the library's. One part is not stable yet, and it is the part with the most hype.

What actually happened

Per the TanStack AI 1.0 release post:

• 23 provider adapters, plus an openaiCompatible adapter for any OpenAI-style API. The alpha had four.
• UI support for React, Vue, Solid, Svelte, Preact, Angular and others, and 11 sandbox providers, including E2B and Railway.
• New since the release candidate: MCP server support with tool filters and approvals, a unified reasoning option, automatic prompt caching, cross-model history replay, and fakeText() for tests with no network or API key.
• Breaking changes: conversationId becomes threadId and is now required on client and server. addToolApprovalResponse() is gone in favor of an interrupts API. The old React, Solid and Vue UI packages have no 1.0 release. Client and server must move to 1.0 together, and stored chats need their own migration that type checks won't catch.
• The agent harness is experimental. @tanstack/ai-harness ships as 0.x, and the post says those APIs can change in a minor release. It also says the harness permissions plugin is not a sandbox: its tools run on the host.

Why TanStack AI 1.0 matters for your business

Most AI features we build for clients sit on one provider's SDK. That is fine until pricing changes, a model is deprecated, or a cheaper one ships. A stable, provider-neutral core makes the model a config value. Cross-model history replay means a conversation can switch providers mid-thread.

Our rules for adopting it:

Build on the 1.0 core. Chat, tools and provider adapters are the stable surface. That is where your product logic should touch the library.

Wall off the harness. If you use the agent harness, wrap it behind your own interface so a minor release can't break your app. Pin the exact version.

Don't confuse permissions with isolation. TanStack says it plainly: the permissions plugin runs tools on your host. If an agent can execute code, use one of the real sandbox providers.

Write the tests first. fakeText() lets you test tool flows in CI for free. Use it before you migrate.

Sources: TanStack: TanStack AI 1.0.