iOS 27 Code Points to a Modular Siri That Can Delegate to Claude—or Swap Its AI Backend

The most important finding in the reported iOS 27 and macOS Golden Gate code is not that users can now replace Siri with Claude or another chatbot. They cannot, at least not through a public Apple setting or developer API. Rather, the private frameworks suggest Apple has built Siri as an orchestration layer: one component can interpret a request, while Siri preserves the interface and carries out Apple-controlled system actions. 36

That design reportedly supports two very different levels of outside-model involvement.

1. Model Delegation: a model interprets the request, Siri performs the action

The narrower mechanism is called Model Delegation and is reportedly built on App Intents. In a demonstration cited by multiple reports, Claude appeared in macOS’s contextual “Ask…” menu alongside the existing ChatGPT-style handoff. The user asked Claude to set a reminder; Claude interpreted the natural-language request, and Siri then created the reminder in Apple’s Reminders app. 36

Conceptually, the flow is:

1. The user chooses a model from Siri’s “Ask…” menu.
2. The outside model turns the request into an understood intent.
3. Siri receives that interpretation.
4. Siri invokes the relevant Apple system action or app action.

This matters because the model is not necessarily being given free-standing control of the device. In the reported example, Siri remains the actor that performs the privileged action. That is a more bounded integration than simply making an external chatbot the operating system’s default assistant. 315

2. Inference Providing: an outside model could become Siri’s reasoning engine

A second reported mechanism goes much further. The private Inference Providing protocol in Model Manager Services was demonstrated with GPT-5.6 Terra in place of Apple’s server-side Siri model. According to the original demonstration, the provider received Siri’s native planner prompt and tool definitions, could make calls to system-action tools, and could receive tool results that include personal data. The answer could then be returned through Siri’s UI and voice, including interactive elements. 35

In this model, Siri would still provide the user-facing shell—voice, presentation, and the Apple-defined tool layer—but the third-party provider would do much of the conversational planning and reasoning. That is qualitatively different from a one-off chatbot handoff. 335

Why the two paths have different privacy implications

The architecture implies a meaningful boundary between delegating interpretation and delegating planning.

With Model Delegation, an external model may be asked to understand a selected prompt, while Siri can remain responsible for access to system actions and personal context. With an inference provider, the external model may need Siri’s prompts, tool schemas, and the results returned by those tools in order to plan a multi-step task. Those results can contain personal data, according to the demonstration. 35

That does not establish unrestricted device access for a model provider. The reported setup still routes actions through Apple-defined tools and permissions. But it would create a substantially broader trust relationship than today’s optional chatbot handoff: the provider could participate in the reasoning loop that decides which personal-data tools to use and how to combine their results. 335

Can users install Claude, Gemini, or another model in Siri now?

No public evidence in the supplied reporting shows that users can install an arbitrary Siri inference provider or that developers can ship Model Delegation extensions today. The relevant frameworks and entitlements are private. Reports say the macOS Golden Gate release candidate’s “Ask…” integration was limited to ChatGPT, while the demonstrated Claude path was not a public product release. 46

So the discovery should be read as evidence of internal capability and design direction, not as an announced marketplace for Siri models.

What about the reported WWDC 2026 Extensions framework?

The record is inconsistent. Several secondary reports claimed Apple had announced an Extensions framework through which users could choose Claude, ChatGPT, Gemini, or Grok for Siri and other Apple Intelligence features. 1923

But another report explicitly says Apple did not announce that framework and that ChatGPT remained the only announced Siri handoff. 18 In light of that conflict—and the lack of a confirmed public entitlement in the code-discovery reporting—it is not reliable to describe a broad, public multi-model Siri Extensions program as established fact.

Did the EU Digital Markets Act drive this design?

It is plausible that interoperability and user-choice pressures, including the EU’s Digital Markets Act (DMA), made a modular model architecture strategically valuable. But the discovered code does not prove the DMA caused Model Delegation or the inference-provider protocol, and Apple has not publicly attributed the design to the DMA.

The situation is also more complicated than a simple “DMA forced Siri open” narrative. Apple said Siri AI would not ship on iOS and iPadOS in the EU at the iOS 27 launch because of DMA-related concerns, with no availability timeline at that point. 45 That makes the code a signal of technical optionality, not proof of a settled regulatory strategy.

What this means for Siri’s future

Siri AI’s reported redesign is conversational, capable of using personal context, able to act in apps, and aware of what is on screen. 3743 The private-model hooks make more sense in that context: a sophisticated assistant needs a stable interface, a controlled action layer, and the flexibility to use different reasoning systems behind them.

If Apple ever exposes these capabilities, it could improve, regionalize, or change the models behind Siri without replacing Siri’s voice, interface, app-action infrastructure, or permission model. But that outcome remains speculative. For now, the strongest conclusion is narrower: Apple appears to have engineered Siri so that an outside model can either assist with an individual request or, under a much more sensitive arrangement, supply the server-side reasoning behind Siri itself. 335