TanStack AI 1.0: An Agent Toolkit Without Lock-In
1.0 landed 23 October 2026: all 76 packages stable, 23 provider adapters, nine UI frameworks, 11 sandbox providers and support for coding agents like Claude Code and Codex — plus subagents, Portable Agent Skills, compaction and computer use, with only the agent harness still experimental
Tool Interface
Interactive tool will be available soon
Features
- ✓ Adapters galore: 23 provider adapters plus openaiCompatible, so any OpenAI-style API fits and swapping models does not mean rewriting business logic
- ✓ Every UI framework: packages ship for React, Vue, Solid, Svelte, Preact, Angular, Octane, Remix and vanilla JS
- ✓ Coding agents in sandboxes: supports Claude Code, Codex, OpenCode, Grok Build and any ACP agent, with 11 sandbox providers
- ✓ Subagents and reusable skills: chat() and useChat hand turns to named child agents such as researchers or writers, and any tool-calling model can load SKILL.md skills on demand
- ✓ Production claims backed by tests: 943 deterministic end-to-end tests across 17 adapters on every pull request, plus compaction and computer use
How to Use
- Move the whole version line together: a 1.0 client needs a 1.0 server, so upgrade every @tanstack/ai* package at once rather than piecemeal
- Check the list: every API deprecated in 0.x is gone in 1.0, but replacements already shipped in 0.x — if you followed the warnings you are most of the way there
- Prove one minimal case through your provider adapter and sandbox path, in local and staging, before migrating real workloads
- If you want the agent harness, go in knowing its APIs are explicitly experimental and may change in minor releases, so keep it out of core paths
FAQ
What is TanStack AI 1.0?
TanStack's open-source AI and agent framework, which reached 1.0 on 23 October 2026 with all 76 packages going stable together. Per the official post, less than a year earlier it was one chat() function and four providers; it now covers chat, tools, media, MCP, persistence, sandboxes and agents, with 23 provider adapters and eight UI frameworks, built by more than 140 contributors. Read it at https://tanstack.com/blog/tanstack-ai-1-0
What breaks between 0.x and 1.0?
Officially there are no new concepts in 1.0: every API deprecated in 0.x was removed, and each replacement already shipped during 0.x. If you kept up with deprecation warnings you are most of the way there. The one hard rule is that you must upgrade the entire version line together — a 1.0 client needs a 1.0 server.
Why not just use one vendor's SDK?
Because the pitch is no lock-in. TanStack frames 1.0 as a formal commitment to the APIs, stressing that your provider, framework, sandbox and the wire protocol between them remain your choice. The payoff shows up when you swap models, change sandboxes or embed agents into editors — business logic does not have to move with it. For scale: four adapters at alpha, 23 now.
Is the testing enough to trust?
The numbers are checkable: 265 deterministic end-to-end tests across 10 providers on every pull request at beta, and 943 across 17 adapters on the 1.0 branch. Publishing test scale is rare and a good sign, but high end-to-end coverage does not make your own integration safe — validate locally and in staging anyway.
Is the agent harness usable yet?
Usable, but not a stable API. The 1.0 announcement singles it out as the one exception: the agent harness APIs are experimental. In practice that means experimenting and evaluating is fine, while putting it on a core path with availability promises is not, and you should expect changes across minor releases.
Which new capabilities should I try first?
The post's 'what landed since the RC' section names: subagents (hand a turn to a named child agent with typed input and limits), decide() for typed choice, score and boolean decisions, Portable Agent Skills (any tool-calling model loading SKILL.md on demand), compaction to stay under the context limit, and computer use running the same code on OpenAI and Anthropic. Start with whichever gap hurts most.