Siri AI opens to third-party apps: your app is now a tool for an AI agent

On November 3, 2026, alongside the iOS 27.3, iPadOS 27.3, macOS 27.3 and visionOS 27.3 betas, Apple released a Siri AI Developer Preview. The short version: any developer can now connect their app's content and actions to the new Siri. That may sound like just another Siri API update, but the documentation tells a bigger story. Apple is turning every app on the iPhone into a tool a language model can call, and it is doing so with concepts the agent world has settled on over the last couple of years.

What was announced?

From Apple's developer news post:

• Any app can expose its content and custom actions to Siri AI through the App Intents framework. When a user asks for something in their own words, Siri can run the action you built: checking a delivery, booking a table, placing an order.
• For complex requests there is a new mechanism called skills, where you describe how Siri should combine several actions into one experience.
• Siri AI can now surface any entity in the Spotlight semantic index, not only those matching Apple's predefined schemas.
• Shipping to customers is a few months out; App Store submissions open in the spring.

One note relevant to this blog's Turkish readers: Apple says Siri AI currently supports English, French, Japanese, Korean, Portuguese and Spanish, and will soon add Turkish, Traditional Chinese, Norwegian and Vietnamese. No date was given, but for Turkish-language apps this integration should move from "someday" to the actual roadmap.

How does it work?

Intents become tools

The core change is the @AppIntent(.tool) macro. It adds ToolIntent conformance to an intent and makes it a tool Siri can call. A short example in my own words:

@AppIntent(.tool)
struct TrackOrder: AppIntent {
    static let title: LocalizedStringResource = "Track Order"
    static let description = IntentDescription("Returns the shipping status of an order.")

    @Parameter(title: "Order", description: "The order to track")
    var order: OrderEntity

    func perform() async throws -> some IntentResult { /* ... */ }
}

The important bits are title and description. Apple states plainly that Siri relies on them to decide when to use an intent and how to tell similar ones apart. These strings are no longer just labels in the Shortcuts app; they are tool definitions read by an LLM. If you've done function calling with the OpenAI or Anthropic APIs, this will feel familiar: a sloppy description means the model picks the wrong tool.

The docs also draw a line: if an intent already conforms to a schema in a domain Apple supports (messaging, for example), you don't mark it as a tool; Siri runs it via the schema. And intents that don't perform a real action in perform(), such as widget configuration intents, can't be tools.

Entities and authentication

For Siri to find your content, entities need to be indexed in Spotlight. With the new @AppEntity(.generable), the model can fill in a transient entity, like a flight search, straight from the user's sentence. On the security side you can set authentication policies per property: require local device authentication before showing an account balance, or hide a field from Siri entirely with visibility: .internal. When you're opening your app's data to a language model, deliberately reviewing these defaults is the most important part of the job.

SKILL.md: a familiar format from the agent world

The most interesting part, in my view, is app skills. A skill is a SKILL.md file in its own folder inside a Skills.skillsbundle added to your Xcode project. It has YAML frontmatter (name and a one-line description) and a free-form Markdown body of up to 10,000 characters. Apple describes it as a subset of the Agent Skills format:

---
name: order-return
description: Walks the user through returning one of their orders.
---
First use the order search tool to find the right order. If there are
several matches, ask the user. Don't call the return tool before you
know the reason for the return.

Siri reads only the description to decide whether a skill matches, and loads the body only when it does. Apple adds an important caveat: skills don't enforce deterministic behavior, and Siri doesn't necessarily run the listed tools in order. This is guidance for a model, not a workflow engine. Critical business rules, like confirming before a payment, still belong in the intent's own code.

Why it matters

Until now, Siri integration was rarely worth the effort for most teams because of fixed phrases and narrow domains. In the new model, a model discovers and calls your app's features based on what users say in natural language. A few practical consequences:

• Discoverability changes. Users can reach your features without opening the app, by naming it or while it's in the foreground. The docs also describe cross-app actions where content found in one app is handed to another.
• Tool design becomes a product skill. Which actions you expose, how you name them and how you describe them are now product decisions. Teams already working with MCP or function calling on the backend can bring that experience straight to mobile.
• Testing gets harder. Model behavior is probabilistic. Apple offers a Siri Integration Advisor in Xcode and a beta Siri Evaluations framework that runs scripted conversations and checks what Siri does.

Limitations

• This is a Developer Preview. Apple explicitly says APIs and Siri behavior may change, and terms and usage policies haven't been published yet.
• The preview only works on Apple Intelligence devices, with development-signed builds and internal TestFlight testers; external testers and App Store builds are excluded. Apple recommends testing on a device, not the Simulator.
• Siri AI is currently unavailable in the EU on iOS, iPadOS and watchOS, and not available to users under 13. Features relying on server-side models are subject to daily usage limits.

What to do now

You can't ship this to production yet, but there's useful work to do before the App Store opens in the spring:

1. Inventory your existing App Intents and decide which ones perform a real action worth exposing as a tool.
2. Rewrite intent, parameter and entity descriptions as if an LLM will read them, because it will.
3. Review authentication policies for sensitive data.
4. Write an experimental SKILL.md for one or two of your most complex flows and measure it with Siri Evaluations.

Apple's move shows that the tool and skill concepts that matured on the server side of the agent ecosystem are now moving down into the operating system. Teams that design their apps for this model early will have a real head start, especially once Turkish support arrives.