Apple made AI visibility mandatory and AI integration provider-agnostic. What changed, why it rhymes with App Tracking Transparency, and what to ship before iOS 27 lands in September.
TL;DR
- App Intents is now the only way Siri reaches your app. No App Intents means your app is invisible to the new Siri.
- Foundation Models added a model abstraction layer. Swap Apple's on-device model, Gemini, or Claude with one line of code.
- The framework now ships a Python SDK, runs on Linux, and accepts image input from third-party apps.
- The EU and China get the developer APIs but not the consumer Siri AI features at launch.
At WWDC 2026, Apple turned two things into platform defaults for mobile builders. First, AI visibility: App Intents is now the only way Siri can reach into your app, and SiriKit is on a deprecation clock. Skip it, and your app disappears from the assistant on iOS 27. Second, AI integration: the Foundation Models framework added a model abstraction layer, a Python SDK, Linux support, and image input, so adding on-device or multi-provider AI is a small, portable change now instead of a bet on Apple's roadmap.
The habit that made this keynote boring (in a good way)
Years ago I kept a browser tab pinned to the Apple developer release notes. When Apple shipped App Tracking Transparency and made the data-sharing prompt mandatory, it broke analytics and attribution across half the apps I knew. You could not measure anything without explicit user consent, and teams scrambled for months to catch up. I had the consent flow wired into my React Native apps the week it went mandatory, because I had been watching it coming for a long time before that.
That habit, watching what Apple makes mandatory and shipping it early, is most of the game on this platform. It is also why yesterday's keynote did not surprise me. The mandatory shift this year is not a tracking prompt. It is whether AI can see your app, and whether your app can use AI without betting the whole architecture on one vendor.
Here is what mattered, filtered for people building mobile products. The React Native specifics are at the bottom.
Why your app is invisible to Siri without App Intents
I had the question before yesterday: how do you make your app visible to the AI tools running inside the phone? I half expected the answer to be an MCP server exposing your backend data, or something equally involved. App Intents was the answer I expected, just not as direct as Apple made it this time.
Siri AI runs on App Intents. SiriKit is deprecated as of WWDC 2026 with compile-time warnings, and App Intents is the only supported way Siri can call into a third-party app. If your app does not expose its core actions as intents, the new Siri cannot see or trigger any of them.
This is the App Tracking Transparency moment repeating. Back then the mandatory thing was a consent prompt. This time, it is a machine-readable description of what your app does. The assistant became the front door on iOS 27, and App Intents is the only key that fits the lock.
App Intents 2.0 added richer entity types, streaming responses for long-running actions, and multi-turn conversational follow-ups. If you ship a consumer app and your users touch Siri, this is the single highest-priority addition to your project before September. I am treating it the way I treated the tracking consent flow in 2021: not optional, and better done early.
This is the work I am doing right now in Morrow Self, the habit app inside my AI Mobile Launcher. Morrow Self sits at the center of someone's daily routine, so making it reachable through Siri is exactly the kind of integration that earns its place: add a habit, check today's progress, log a streak, all by voice, without opening the app. I am wiring its core actions up as App Intents ahead of iOS 27 so the assistant can drive it directly.
Why App Intents need generative UI, and why that part is not solved yet
App Intents answer one half of the question: the assistant can now call your app's actions. They do not answer the other half, which is what the user sees when it does. An intent can hand back a small SwiftUI snippet, but the moment the response is richer than a confirmation, a chart, a multi-step form, a list the user can act on, you are back to hand-building a fixed screen for every possible response.
That is the generative UI gap. On the web the pattern is settled: the model picks from a set of registered components and fills them in, so the interface adapts to the request instead of being hard-coded for it. On mobile, that pattern barely exists in any packaged form. App Intents gives you a typed surface of actions the assistant can choose between. Generative UI would give you a typed surface of components it can compose into a response. Apple shipped the first half this year. The second half is still an open problem, and nobody has a clean answer for it yet.
I care about this because it is the real product question hiding behind "make your app usable by AI." Being callable is table stakes. Being able to answer with the right interface, generated for that moment instead of pre-built for it, is where the experience is won. It is also why I am building Wire RN: a way to register React Native components an assistant can pick between, the same way App Intents lets it pick between actions. The intent layer is standardized now. The UI layer is the part still up for grabs.
Apple just made AI providers swappable in one line
Foundation Models in iOS 27 introduced a LanguageModel protocol (docs here). Apple's on-device model, Google Gemini, and Anthropic Claude all implement it. You write your session logic once, and switching providers is a one-line change.
Until yesterday, building on Foundation Models meant committing your AI architecture to Apple's model and Apple's release schedule. Most teams I talk to refused that bet and built on the OpenAI API instead, eating the cost and the privacy trade-offs. The abstraction layer removes the reason to refuse.
The practical shape is hybrid routing. Short, private, or simple calls go to the on-device model, which is free and works offline. Complex reasoning routes to Claude or Gemini through the same API. You pay only for the calls that need frontier quality. Production teams have built this by hand for two years, and Apple just made it a one-day job.
Foundation Models now runs on Linux and speaks Python
Apple shipped a Python SDK for Foundation Models and made the framework run on Linux. An Apple Intelligence-adjacent runtime is now callable from a Python script on a Linux server.
That sentence would have been absurd a year ago. It means you can prototype mobile AI features without an iOS device, run the same evaluation suite on CI and on-device, and target Apple's model from a Python agent framework next to Claude and Gemini. The cost of building for the Apple stack just fell to roughly what it costs to build for the OpenAI API.
The on-device model also accepts image input from third-party apps now. Any app doing photo analysis, OCR, or scene understanding can drop its bundled 1 to 2 GB vision model and call the system API instead. Cal AI, the photo-to-calorie app that hit 40 million dollars in revenue before MyFitnessPal acquired it, is built on exactly this kind of vision call. After yesterday, that capability is free and on-device for everyone.
What the Apple and Google deal admits
Apple Foundation Models on Cloud is the frontier-quality tier, and it runs on Google's infrastructure. AFM Cloud Pro runs on Nvidia Blackwell B200 chips inside Google Cloud, under a partnership that reportedly costs Apple around 1 billion dollars a year for a custom Gemini model. Apple's contract bars Google from training future models on Siri queries.
Apple framed this as a privacy decision. The simpler read is that Apple cannot build frontier LLMs alone yet, so it rented Google's. For builders, the upside is that the cloud tier gives you Gemini-comparable quality through Apple's API surface, wrapped in Apple's privacy contract. If you need frontier reasoning in a regulated context, that is a real option now.
The EU and China caveat, which is personal for me
Siri AI consumer features are not launching in the EU or China at WWDC 2026, citing regulatory constraints. If you ship consumer apps in those markets, your users do not get the new Siri at launch.
The nuance that matters: the Foundation Models developer API works everywhere. Your apps can call the on-device model, the cloud model, and the multi-provider abstraction in every region. The gap is on the consumer side, not the developer side. I am in Berlin, so the people around me will not see Siri AI on day one. The apps I build can still use the full framework today, which is the part worth acting on.
What to do before iOS 27 ships
- Audit your iOS app for App Intents coverage. If you have none, your app is invisible to the new Siri. Plan the migration off SiriKit now, not in August.
- Wire the on-device Foundation Models into one real feature, behind an adapter, so you can swap providers later without a rewrite. Start with structured extraction or classification, not chat.
- If you ship native iOS, connect Claude Code or OpenAI Codex to your Xcode MCP server. Xcode 26.3 already exposes build, test, and diagnostics over MCP. This is the workflow change with the highest payoff from the keynote.
- If you build an AI product, apply for the Extensions developer beta. Apple opened Siri Extensions so users can pick Claude, ChatGPT, or Gemini as their assistant provider. That is a distribution channel on 2 billion devices, and the public launch is about three months out.
For React Native developers
Most of this translates to React Native, with one caveat and one opportunity.
The caveat: Foundation Models is a Swift framework, so calling it from RN needs a bridge. The cleanest path today is @react-native-ai/apple from Callstack, which already wraps the framework in a Vercel AI SDK-compatible interface. It was built against the iOS 26 version, so the iOS 27 additions (model abstraction, multimodal, fine-tuning) will need bridge updates before you can use them from RN. Watch the repo.
The opportunity: the multi-provider story Apple shipped on native is the same one @react-native-ai was already building toward. Register providers, swap with one line, same API. The two are now describing the same architecture from opposite ends, so the RN side should catch up fast.
For App Intents, you expose them through Swift in the native side of the project, and there are community packages that reduce the boilerplate. For MCP, you do not need anything RN-specific: the Xcode integration works at the Xcode level, so point your agent at the workspace and it works. And Wire RN, the open-source generative UI SDK I am building at getwireai.com, will track the Foundation Models multi-provider API as it stabilizes.
FAQ
Do I have to rewrite my app to support App Intents?
No. App Intents are added alongside your existing UI, not instead of it. You declare your app's core actions as intents in Swift, and Siri, Spotlight, Shortcuts, and the Action button can all call them. Your existing screens keep working exactly as they do now.
Is the on-device model good enough to replace cloud API calls?
For some tasks, yes. The on-device model handles classification, structured extraction, summarization, and simple tool calling well, all for free and offline. For complex reasoning, long context, or frontier-quality generation, you still route to the cloud. The point of the new abstraction layer is that you can do both through one API.
Can React Native apps use Apple Foundation Models?
Yes, through a bridge. The @react-native-ai/apple package from Callstack wraps the framework in a Vercel AI SDK-compatible interface. It currently targets the iOS 26 version of Foundation Models, so the new iOS 27 features will need bridge updates before they are available from React Native.
Does Siri AI work in the EU?
The consumer Siri AI features are not available in the EU or China at launch, due to regulatory constraints. The Foundation Models developer API, however, works in every region. Your app can use on-device and cloud models in the EU even though end users will not get the new consumer Siri yet.
Is App Intents only useful for Siri?
No. App Intents power Siri, Spotlight search, the Shortcuts app, widgets, and the Action button. Adopting them makes your app reachable across all of those surfaces, not just the assistant. That breadth is why it is worth doing even before iOS 27 ships.
What I'm shipping this week
I am going back through my own apps with the same checklist I just gave you, starting with App Intents coverage in Morrow Self, because I would rather migrate off SiriKit now than during the iOS 27 beta crunch. On the Wire RN side, the model abstraction announcement lines up almost exactly with the provider-swap design I was already building, so I am revisiting that interface to match the LanguageModel shape Apple just standardized.
If you want help shipping any of this on a React Native or native iOS stack, including App Intents migration or a multi-provider AI setup, that is what CasaInnov does.
I write Code Meet AI on AI-first mobile development. Next issue goes deeper on the Future of Mobile, the foundations worth building on once on-device AI and generative UI become table stakes.
Top comments (0)