TanStack AI 1.0: Stable SDK, Upgrade and Agent Limits

TanStack released TanStack AI 1.0 on October 23, 2026, turning its provider-agnostic TypeScript AI toolkit into a stable option for production chat, tool-calling and multimodal applications. The release covers 76 stable packages, 23 provider adapters and a range of UI frameworks. The headline needs one important qualification: its new durable agent harness is explicitly experimental and still ships as 0.x packages.

What is stable in TanStack AI 1.0?

The stable SDK combines server-side chat and model adapters with framework integrations, typed tool definitions, streaming responses, structured outputs, media generation, MCP connections and sandbox options. TanStack says its 1.0 branch runs 943 deterministic end-to-end tests across 17 provider adapters. These are project-reported test counts, not independent reliability or performance benchmarks.

For an application team, the practical attraction is choice: a React, Vue or Svelte interface can communicate with a server using a model adapter without committing the entire product to one model provider. That flexibility does not make provider pricing, model capabilities, authentication rules or tool behavior identical. Test the exact providers and workloads your application needs.

Upgrading from 0.x requires coordinated changes

TanStack describes 1.0 as a compatibility commitment, not a drop-in upgrade for every 0.x app. Deprecated APIs were removed. A 1.0 client must communicate with a 1.0 server, so update all @tanstack/ai packages together. The official migration documentation identifies renamed fields and AG-UI wire-format changes that can break custom endpoints if only the frontend is upgraded.

1. Inventory every @tanstack/ai package in the client and server; plan one coordinated version upgrade rather than mixing 0.x and 1.0.
2. Run the official AG-UI codemod in a branch and review its changes before merging.
3. Replace the old body option with forwardedProps and the old conversationId option with threadId. Provide a stable threadId on both client and server.
4. Inspect custom API routes: older body.data reads must become forwardedProps reads. Explicitly allowlist client-controlled fields instead of spreading them into model configuration.
5. Replace deprecated framework UI packages with the supported framework package UI entry points, where relevant.
6. Run type checks, test tool approvals and streaming, and follow the saved-data migration guidance for persisted chats, runs or files before deploying.

Where the new agent harness fits

The harness is designed for agents that need more than a single chat request: durable sessions, user steering while a turn runs, background work, tool approvals, and sessions that can resume after a restart when backed by persistent storage. TanStack documents one harness definition that can be served through a web app, command-line interface, ACP or MCP. These are developer building blocks, not a hosted autonomous agent service that works without deployment, storage or operational safeguards.

A practical evaluation before adopting it

Start with one real workflow: a chat interface that calls a read-only tool, requests approval before a write, and preserves its conversation across a reload. Measure successful completions, tool-call errors, latency and human review effort on your own data. If provider switching matters, repeat the same test with a second supported adapter. Keep API keys on the server unless you intentionally implement a documented bring-your-own-key flow.

The October 23 release is meaningful for TypeScript teams because the core API now has a stable major version and a documented migration path. Its experimental harness broadens the possibilities for longer-running agents, but it should be assessed separately from the production-ready core.

Questions people ask