The most consequential thing in Apple's newest operating systems is not a feature users can see. Buried in iOS 27 and macOS 27 Golden Gate are private frameworks that let an outside language model stand in for Siri's own server-side planner — receiving Apple's internal prompts, calling Apple's system tools, and touching the personal data those tools return.
Key takeaways
- An Inference Provider protocol inside Model Manager Services can substitute a model such as GPT-5.6 Terra for Apple's server-side Siri model, passing it Siri's planner prompt and tool definitions.
- A lighter-weight Model Delegation API in App Intents registers a model like Claude as a Siri extension, alongside the existing built-in ChatGPT integration.
- Neither mechanism is live: the entitlement is closed to third-party developers, and the macOS 27 release candidate still offers only ChatGPT in its "Ask…" menu.
What the inference provider protocol actually does
A developer working under the handle pdfu surfaced the frameworks and published video demonstrations of both. The second one is the striking one. Rather than sitting beside Siri, the substituted model takes Siri's place in the pipeline: it inherits the native planner prompt Apple writes, along with the tool definitions that describe what the assistant is permitted to do on the user's behalf.
That inheritance is what makes the swap complete rather than cosmetic. The external model can issue tool calls that perform real system actions, take delivery of whatever personal data comes back, and hand Siri a finished answer that Siri then speaks in its own voice through its own interface. One clip walks a single request across four surfaces — locate emails on a topic, summarise their contents and action points, then message a contact through Messages — and the corresponding request appears in OpenAI's platform console, a reminder that the reasoning has left the device entirely.
The shallower path: Siri extensions
Model Delegation, the other mechanism, is closer to what Apple already ships. It lets a model register through App Intents and appear the way the built-in ChatGPT extension does. In a macOS clip, Claude is selected from an "Ask…" contextual menu in the "Search or Ask" bar and asked to set a reminder; it reads the request, then defers to Siri for the part that touches Apple's own Reminders app.
The handoff runs both directions. Where Siri hits its own ceiling — a request to build a CSV file, in another clip — Claude produces the result inside the Siri conversation window. The design treats the two assistants as complementary rather than competing, with Apple retaining control of anything that reaches into system state.
Why the plumbing exists at all
Regulation is the likeliest explanation. Under the European Union's Digital Markets Act, Apple must give third parties effective access to the iOS hardware and software capabilities it reserves for its own services, and the European Commission has stated that the obligation reaches Siri specifically. Engineering interoperability into the architecture up front costs less than bolting it on against a compliance clock.
The same logic bounds the upside. 9to5Mac raised the prospect that the capability ships only in the EU, mirroring how alternative app marketplaces and browser choice screens were handled. Nothing in the recovered code settles whether users elsewhere ever see it.
Outlook
None of this is a shipping product. Apple has said nothing publicly, the entitlement remains closed, and Claude does not appear in the release candidate that reached users this week. The timing is awkward in its own right: Apple only just shipped Siri AI in iOS 27 as an English-only beta, meaning the model these frameworks could displace is not finished either. If the entitlement ever opens, the AI assistant slot on the iPhone stops being Apple's by default and becomes something the company has to win.
FAQ
Can I replace Siri with Claude or ChatGPT right now?
No. The frameworks ship inside iOS 27 and macOS 27 Golden Gate but are not switched on for users, and Apple has not granted the model delegation entitlement to outside developers. The "Ask…" menu in the macOS 27 release candidate still exposes only the existing ChatGPT extension.
What is the difference between the two mechanisms found in the code?
Model Delegation registers an outside model as a Siri extension via App Intents, so the model interprets requests but returns system actions to Siri for execution. The Inference Provider protocol goes further, replacing Apple's server-side Siri model and granting the external model the planner prompt, the tool definitions and the personal data those tools return.
Would this feature be available outside the European Union?
That remains unclear. The Digital Markets Act is the most credible reason Apple built the capability, and the European Commission has named Siri in its interoperability requirements. Apple has confined comparable accommodations to the EU before, so a region-locked rollout would not be surprising.
