What changed
On October 23, 2026, the TanStack maintainers released TanStack AI 1.0, making its core chat, tool, media, MCP and streaming interfaces stable. The project reports 23 provider adapters and support for several UI frameworks, plus sandbox and coding-agent integrations. The release also introduces an experimental long-running agent harness, packaged separately at 0.x rather than covered by the 1.0 stability commitment. Migrating from 0.x requires coordinated 1.0 client and server packages, replacing deprecated APIs, updating streaming completion events and handling stored conversation data. TanStack says its core tests use deterministic model fixtures rather than live provider calls.
Why it matters
Teams building AI features into existing web applications can choose a framework-neutral SDK rather than tie their UI and provider to one proprietary stack. Stable core APIs reduce the risk of surprise changes, but the upgrade isn't a zero-work dependency bump: existing clients and servers must agree on protocol versions and saved state may need migration. The more ambitious background agent features are explicitly experimental, so production builders should separate the stable core from evolving harness internals. A small open-source project can change development practice without having a large commercial sponsor.
Stable means core APIs, not every package
TanStack's maintainers say the 1.0 contract applies to the core chat, tools, media and integration APIs. The new @tanstack/ai-harness and CLI remain experimental 0.x packages, as do relevant harness entry points elsewhere. Avoid treating a stable package label as proof that background agents, scheduling and persistence APIs are fixed.
Provider and UI independence is the central design
The project reports 23 provider adapters plus an OpenAI-compatible option, and integrations spanning React, Vue, Svelte, Solid and other UI layers. It also exposes MCP tools, supports provider switching and supplies deterministic test helpers. Those are upstream-reported capabilities; a production deployment should verify the exact providers and modalities it needs.
The 0.x to 1.0 migration can break connected clients
The official upgrade guide says all stable @tanstack/ai* packages should move together because a 1.0 client expects a 1.0 server. Deprecated APIs were removed; the project provides a codemod, but developers must supply thread IDs, handle tool approvals through interrupts, replace certain UI hooks and stop emitting the old [DONE] streaming marker in favour of RUN_FINISHED. Passing TypeScript checks does not migrate saved conversations, runs or files.
The harness targets sessions that outlive requests
The new defineHarness/createHarnessHost approach lets applications model persistent sessions, subagents, plugins, background workers, tool permissions and multiple transports including HTTP, ACP and MCP. These features may enable coding agents and support bots to survive deployments, but their APIs remain experimental and may change in minor versions. Operational safety still depends on persistence backends, authentication and idempotent external actions.
A small maintainer team needs practical validation
TanStack credits three maintainers and 143 contributors. Its deterministic provider-adapter tests reduce regressions but are not live tests against every model endpoint or sandbox. Builders should stage upgrades with representative tool calls, provider-specific edge cases, streaming reconnects and persisted conversation replay before using the new version in production.