TanStack AI 1.0: Still Your AI, Still Your Way

TanStack AI has reached 1.0, and the headline is less about shiny new tricks than about promises kept. The core APIs are now stable for production, while the new agent harness lands as experimental on purpose. The project’s pitch hasn’t changed: pick your provider, your framework, your sandbox, and even the wire protocol, without getting trapped in one stack.

That promise has grown a long way from the alpha. What began as a chat() function with four providers is now a broader toolkit covering chat, tools, media, MCP, persistence, sandboxes, and agents. The provider count is up to 23, the UI framework list now stretches to eight, and more than 140 people have contributed along the way.

The release is also unusually tidy for an AI library. If someone kept up with the deprecation warnings, there are no new concepts to learn here. The old 0.x APIs that were marked deprecated are gone, but their replacements already shipped earlier. The one catch is practical rather than dramatic: upgrade the stable @tanstack/ai* packages together, because a 1.0 client needs a 1.0 server.

The rougher edges were already sanded down in the run-up to 1.0. Since the RC, TanStack AI says it has added subagents, typed decisions, portable skills, compaction, computer use, reasoning options that work across providers, prompt caching, history that can move between models, and a unified Files API. It also added new adapters for Cloudflare Workers AI and AI Gateway, Lovable AI Gateway, Azure OpenAI, World Labs Marble, TypeSafe, and Ollaya, plus five sandbox providers.

But the most interesting part is still the harness. TanStack is trying to make one session survive restarts, waits, approvals, background work, and even coding agents that keep going while a user does something else. That’s a harder problem than chat boxes, and probably the more useful one too. The stable core gives the project a foundation; the experimental harness is the bet that AI apps need more than a prompt loop.