TanStack released the alpha of TanStack AI, a toolkit for building AI features without committing to one framework or model provider. It works with OpenAI, Anthropic, Gemini and Ollama, and ships clients for vanilla JavaScript, React and Solid. Details are in the announcement.
My read: it’s a vendor-neutral answer to the Vercel AI SDK. The AI SDK is open source, but when I built tool calls for Inbound, it preferred Vercel AI Gateway credits over the OpenAI key I had configured.
Summary
- Server support covers JavaScript and TypeScript, with Python and PHP in the alpha, though those aren’t on pip or Composer yet.
- A documented protocol lets any backend and transport talk to the client through a connection adapter.
- Tools are isomorphic: define one once, then provide separate server and client implementations with shared types.
- Types are per model, so your editor knows each model’s options and supported modalities.
- Headless chat components for React and Solid come unstyled, in the spirit of TanStack Table and Form.
- It’s an alpha, so expect bugs and rough edges.
What it means for you
- Keep model choice behind one typed interface to avoid provider lock-in.
- Start with the working examples, including TanStack Start, Laravel and FastAPI, before adopting it.
- Use the headless components for full control over chat styling, or look at Assistant UI if you need thread management built in.
- Wait for a stable release before moving production code.
Get the next test the day it’s out.
