TanStack AI 1.1.0: What Changed and How to Upgrade (2026)

@tanstack/ai 1.1.0 shipped on 4 Nov 2026. The verified release record, the cancellation changes that matter, and a safe upgrade path for AI SaaS apps.

Huifer
Huifer
November 4, 202611 min read
Also available in:Deutsch · 中文

Written by Huifer, solo developer and maintainer of TanStack Ship. When npm moved @tanstack/ai to the latest dist-tag on 2026-11-05, I did what I do for every dependency bump in my own templates: I pulled both published tarballs — 1.0.1 and 1.1.0 — and diffed the shipped type declarations instead of trusting a changelog summary. That diff is the backbone of this guide, and it changed what I expected the release to be about.

Verified sources: TanStack · TanStack AI docs · @tanstack/ai 1.1.0 on npm · TanStack AI on GitHub · web.dev · Cloudflare Workers docs

Last updated: 2026-11-05 · Changelog

@tanstack/ai 1.1.0 is the current latest release of TanStack's type-safe TypeScript AI SDK. npm published it at 2026-11-04T14:01:38Z and the latest dist-tag moved on 2026-11-05. It is a minor bump from 1.0.1 (2026-10-29), so semver requires it to stay backwards compatible. The published README is identical to 1.0.1 apart from two sponsor logo URLs, so the useful diff lives in the code — and it says one thing loudly: cancellation now propagates end to end, from the caller into MCP tool discovery, generation activities, realtime token minting, file uploads, and video jobs.

TL;DR

  • Measured: @tanstack/ai 1.1.0 was published to npm at 2026-11-04T14:01:38Z; the latest dist-tag moved on 2026-11-05 (source: npm registry record).
  • Measured: 12 @tanstack/ai* packages published between 13:58:44Z and 14:02:01Z — a 3-minute, 17-second release window.
  • Calculated: 65 of the 644 files in the published tarball changed between 1.0.1 and 1.1.0 — roughly 10% of the package.
  • Measured: the README is byte-identical except for two sponsor image URLs. The docs did not change; the code did.
  • Measured: the API additions are cancellation-shaped — abortSignal and timeout on generation activities, signal on MCP tool discovery, abortSignal on realtime token and connect calls.
  • My call: low-risk bump, but upgrade the whole @tanstack/ai* family in one commit, because the adapters ship separately from the core.

Short answer: @tanstack/ai 1.1.0 is TanStack's type-safe AI SDK, published to npm on 2026-11-04 as the latest tag after 1.0.1. Its measurable change is cancellation plumbing: abortSignal and timeout now flow from chat middleware into MCP discovery, embeddings, realtime tokens, file uploads, and video jobs.

TanStack AI 1.1.0: What Changed and How to Upgrade (2026)

This guide is for the developer with an AI feature in production who now sees @tanstack/ai@1.1.0 in a dependency audit — and for the founder deciding where the AI layer should live. If you are still choosing a starting point, skip to Which AI SaaS template capability this release supports.

Everything below is measured against the two published npm tarballs and the registry metadata, because most release summaries here are written from a changelog that does not exist in the registry payload.

What @tanstack/ai 1.1.0 Actually Ships: The Verified Record

What the npm record proves

The registry entry for 1.1.0 is the authoritative artifact. Here is what it contains, as published:

Registry fieldValue in @tanstack/ai@1.1.0
Version1.1.0 (previous: 1.0.1, published 2026-10-29)
Publish timestamp2026-11-04T14:01:38.341Z
Description"Type-safe TypeScript AI SDK for streaming chat, tool calling, agents, structured outputs, and multimodal generation."
LicenseMIT
EnginesNode >=18
Repositorygit+https://github.com/TanStack/ai.git (directory packages/ai)
Runtime dependencies@ag-ui/core@1.0.0, @tanstack/ai-utils@^1.0.1, @tanstack/ai-event-client@^1.0.1, @cfworker/json-schema@^4.1.1, @standard-schema/spec@^1.1.0, fast-json-patch@^3.1.1, partial-json@^0.1.7
Peer dependency@opentelemetry/api >= 1.9.0 (optional)
Export subpaths., ./byok, ./locks, ./client, ./testing, ./adapters, ./byok/server, ./middlewares, ./middlewares/otel, ./adapter-internals
Package size644 published files; sideEffects: false

Two rows matter most for an upgrade decision. The only dependency ranges that moved are the two internal TanStack packages — @tanstack/ai-utils and @tanstack/ai-event-client, both from ^1.0.0 to ^1.0.1; everything external is unchanged. And the export map did not change at all, so no import path in your app needs rewriting.

What the npm record does not prove

The registry payload contains no changelog body, and the README inside the 1.1.0 tarball is identical to the 1.0.1 README except for two sponsor logo URLs. "What's new" therefore cannot be answered from registry metadata alone — it has to be answered by diffing the published build, which is what I did.

The Actual Change: Cancellation and Timeout Propagation

Diffing the two published tarballs shows 65 changed files out of 644, spread across chat, MCP, middleware, embeddings, files, realtime, and video. The theme is one: an abort signal now reaches the work, instead of stopping the wrapper around it.

Chat middleware now sees one linked abort signal

In 1.0.1 the chat runner kept two separate ideas of "cancelled": the caller's abortController.signal and the middleware's own abort() controller. In 1.1.0 the runner computes a single linkedSignal — the caller signal plus ctx.abort() — and hands that to middleware and to the adapter request. The shipped comment says it plainly: /** The caller signal plus ctx.abort(), only when there is a caller signal. */

For an app that cancels a chat from the UI, this is the difference between "the request stopped" and "the request, the tool loop, and the provider call all stopped". The middleware type declaration was updated to match: signal now aborts when the caller aborts the chat request or a middleware calls abort().

MCP tool discovery receives the run signal

MCP is where an AI app most often hangs. In 1.0.1, MCPToolManager.discover() took no arguments, and each MCP source's tools() call could not be cancelled. In 1.1.0 the type is:

typescript
export interface MCPToolSource {
  tools: (options?: {
    lazy?: boolean
    signal?: AbortSignal
  }) => Promise<Array<AnyServerTool>>
  close: () => Promise<void>
}

and MCPToolManager.discover(signal?: AbortSignal) forwards the chat run's signal into every source's tools() call. The documented contract is explicit that a source which ignores signal still works — this is opt-in cooperation, not a hard requirement.

Generation activities accept abortSignal and timeout

The embed() activity gained two options that did not exist in 1.0.1, and the declaration comments define the semantics precisely:

typescript
export declare function embed<TAdapter extends EmbeddingAdapter<string, any, any, any>>(
  options: EmbedOptions<TAdapter>,
): Promise<EmbeddingResult>
typescript
// inside EmbedOptions in 1.1.0
timeout?: number // Maximum duration of this activity invocation in milliseconds.
                 // No SDK-wide default — choose a value suitable for the provider and job.
                 // Composed with abortSignal; the first abort wins.
abortSignal?: AbortSignal // Caller cancellation signal (request disconnects, job/runtime
                          // cancellation). Composed with timeout into an effective signal
                          // forwarded to the adapter. Request-specific — not stored on global
                          // provider client config.

There is no SDK-wide timeout default. That is a deliberate design decision in the type surface, and it means the bound is yours to set per activity — which is exactly what you want for an embedding job that must not outlive a request.

Realtime, files, and video adapters accept an abort signal

The same shape propagates through the adapter layer:

  • Realtime: RealtimeTokenAdapter.generateToken(options?: { abortSignal?: AbortSignal }), RealtimeTokenOptions.abortSignal, and connect(token, clientTools?, options?: { abortSignal?: AbortSignal }). The client-side connection setup can now be cancelled, not just the token request.
  • Files: upload, get, and delete on FilesAdapter each take an optional { abortSignal }, with the instruction that adapters forward it to the provider request when supported.
  • Video: getVideoStatus and getVideo accept { abortSignal }, and a new optional cancelVideoJob(jobId, options?) lets an adapter ask the provider to stop a job that is still running. The shipped contract states it is best-effort: a rejection is logged, not thrown.

The 2026-11-04 Release Train: Align the Whole AI Family

@tanstack/ai is a monorepo release, and the packages do not share a version number. Here is what I verified on npm for 2026-11-04:

PackageVersion publishedTimestamp (UTC)
@tanstack/ai-anthropic1.1.013:58:44
@tanstack/ai-client1.1.013:58:45
@tanstack/ai-event-client1.0.113:58:50
@tanstack/ai-openrouter1.1.013:59:02
@tanstack/ai-openai1.1.013:59:23
@tanstack/ai-mcp1.1.013:59:50
@tanstack/ai-react1.0.213:59:53
@tanstack/ai-gemini1.1.014:00:14
@tanstack/ai-solid1.0.114:00:30
@tanstack/ai-cloudflare1.1.014:01:37
@tanstack/ai1.1.014:01:38
@tanstack/ai-utils1.0.114:02:01

@tanstack/ai-vue (1.0.1), @tanstack/ai-svelte (1.0.0), and @tanstack/ai-preact (1.0.1) were not part of this train; their latest versions still carry earlier publish timestamps.

Why version alignment matters more than the core bump

Because the framework and adapter packages carry their own versions (ai-react at 1.0.2, ai-utils at 1.0.1), a partial upgrade is entirely possible — and it is the most likely way this release bites you. The core declares @tanstack/ai-utils: ^1.0.1 and @tanstack/ai-event-client: ^1.0.1 as caret ranges, so npm will happily resolve a stale 1.0.0 in your tree if your lockfile pins it, and you would run a 1.1.0 core against 1.0.0 internals with no error telling you so.

Treat the family as one unit: bump every @tanstack/ai* entry in one commit, then reinstall so npm resolves the whole set from the same release window.

Verify It Yourself in Four Commands

This is the whole verification loop, and it takes about a minute.

bash
# 1. What the registry says about 1.1.0
npm view @tanstack/ai@1.1.0 version description license engines peerDependencies

# 2. Who owns the latest tag, and when each version landed
npm view @tanstack/ai dist-tags
npm view @tanstack/ai time --json | grep -E '"1\.(0|1)\.'

# 3. The dependency ranges that actually moved
npm view @tanstack/ai@1.0.1 dependencies
npm view @tanstack/ai@1.1.0 dependencies
bash
# 4. Diff the two published builds instead of trusting a changelog summary
mkdir -p /tmp/ai-diff && cd /tmp/ai-diff
npm pack @tanstack/ai@1.0.1 --silent
npm pack @tanstack/ai@1.1.0 --silent
mkdir -p old new
tar -xzf tanstack-ai-1.0.1.tgz -C old
tar -xzf tanstack-ai-1.1.0.tgz -C new
diff -rq old/package/dist new/package/dist | wc -l
diff -u old/package/dist/esm/activities/chat/mcp/types.d.ts \
        new/package/dist/esm/activities/chat/mcp/types.d.ts

Step 4 is the one that earns its keep: it is how I found the MCP signal addition and the timeout/abortSignal options on embed(), neither of which appears in the README that ships with the package.

Which AI SaaS Template Capability This Release Supports

This is a technical release, so here is the buyer decision it actually informs rather than a generic sales paragraph.

Cancellation plumbing is the difference between an AI feature that costs what you expect and one that quietly bills for work nobody received. When a user closes the tab mid-answer, the provider request, the MCP tool discovery, and the embedding job all keep running unless something cancels them. With 1.1.0, a TanStack Start route can hand its request.signal into the activity, and a cooperative MCP server stops its discovery when the run is cancelled.

That is a template capability, not an application detail. A boilerplate that ships streaming chat, tool calling, structured outputs, and credit metering should also ship cancellation, or every customer re-implements it — which is exactly what the AI SaaS boilerplates I compare are meant to expose.

Concretely, three surfaces are worth checking in any AI SaaS starting point:

  • Streaming chat — chat() plus toServerSentEventsResponse() from the official docs, with the SSE response flushed on a Worker runtime as covered in the Cloudflare Streams API guide.
  • Tool calling and MCP — a toolDefinition() contract you can implement server-side, and an MCP layer that no longer blocks on a cancelled run.
  • Structured outputs — outputSchema with Zod, ArkType, Valibot, or plain JSON Schema, so a parsed result is a typed object rather than a string you regex.

If you want working examples rather than an API reference, the AI Image Studio template ships the image pipeline (generation, edit, upscale, gallery, credits) and the AI Workspace template ships multi-model chat, RAG, web search, MCP tool calls, credits metering, and Stripe billing — all on TanStack Start.

Limitations, Disclosure, and How to Read the Claims

Claim types used here. Measured means read directly from the npm registry metadata or the published tarball. Calculated means arithmetic I performed on those files (the 65-of-644 count, the release-window duration). Estimated means my judgment. User-provided means nothing — no metric here came from a user or a customer.

What I did not do. I did not run this release in a production application. I compared two published tarballs and read the shipped type declarations; I did not benchmark streaming latency, token cost, or cancellation latency, and I claim no improvement in those. The upstream changelog for 1.1.0 was not available to me, so if the maintainers shipped behaviour changes beyond the type surface — a default value change, a runtime fix with no type change — this guide will not show them. Check the TanStack AI GitHub repository for the source diff before you ship.

Disclosure. I maintain TanStack Ship, and the templates linked above are my own products. I have no material connection to any model provider named here, and I make no claim about any provider's pricing, availability, or model lineup.

Conclusion and Next Step

@tanstack/ai 1.1.0 is a well-behaved minor release: the export map is unchanged, the external dependencies are unchanged, the README is effectively unchanged, and the entire API delta is a set of optional abortSignal and timeout parameters that make cancellation reach the work instead of stopping at the wrapper. The risk is not in the API — it is in upgrading only the core and leaving adapter packages a version behind.

Next step: compare AI SaaS boilerplates side by side and check which ones ship streaming chat, tool calling, credits metering, and cancellation wired end to end — then browse the templates and preview the one that matches your product before you write the first prompt handler.

Related reading: Tanstack React Query 5 106 0 Release Guide · Tanstack Redact Webpack Rspack Setup Guide · Tanstack Ecosystem Complete Guide 20260620

FAQ: @tanstack/ai 1.1.0 Upgrade

Is @tanstack/ai 1.1.0 a breaking change?

No. It is a minor version bump from 1.0.1, and semver requires backwards compatibility. I verified that at the artifact level: the export map is identical between 1.0.1 and 1.1.0, the README differs only in two sponsor logo URLs, and every external dependency range is unchanged. The new options are optional, so existing calls compile and run unchanged.

Do I need to upgrade every @tanstack/ai package at the same time?

You should. The family does not share a version number — ai-react moved to 1.0.2, ai-utils to 1.0.1 — and the core depends on @tanstack/ai-utils and @tanstack/ai-event-client with caret ranges. A partial upgrade can leave a 1.1.0 core running against 1.0.0 internals with no error surfacing it.

Does @tanstack/ai 1.1.0 change how I write a streaming chat endpoint?

No. The published README for 1.1.0 still documents the same chat() call and toServerSentEventsResponse() transport as 1.0.1, and the diff shows no change to the chat entry point's type surface. If your endpoint already works, it keeps working. What changed is what happens underneath when a run is cancelled.