Apple Siri Frameworks Expose Model Delegation and Inference Providing Protocols
Internal frameworks in macOS Golden Gate and iOS 27 show Apple engineered Siri to delegate tasks to third-party models like Claude and swap its server reasoning engine via Inference Providing protocols. While Apple has not yet opened the entitlement to developers, the architecture prepares the OS for deep multi-model interoperability.
Impact: High
Why it matters
Rather than locking users into a proprietary reasoning pipeline, Apple's architecture reveals a modular design capable of routing prompts, tool definitions, and system executions directly to frontier models.
TL;DR
- 01Private frameworks in macOS Golden Gate and iOS 27 include Model Delegation and Inference Providing protocols.
- 02Model Delegation allows models like Claude to interpret requests and trigger native Siri actions like Reminders.
- 03Inference Providing allows a model such as GPT-5.6 Terra to fully replace Siri's server planner, receiving tool definitions directly.
- 04The entitlement is not yet publicly opened to third-party developers, and only ChatGPT is active in the macOS Golden Gate RC.
Key facts
- Operating Systems
- macOS Golden Gate, iOS 27
- Protocols Discovered
- Model Delegation, Inference Providing
- Framework Location
- Model Manager Services
- Current Public Availability
- Entitlement not yet open to third-party developers; ChatGPT active in RC
Model Delegation Architecture
Apple has structured its next-generation assistant architecture around modular agent handoffs. Through the Model Delegation protocol uncovered in macOS Golden Gate and iOS 27 private frameworks by researcher pdfu, Siri can function as an orchestration layer rather than a closed pipeline. When invoked via the 'Search or Ask' bar, an external model such as Claude receives the user query, parses intent, and hands execution back to local macOS and iOS APIs via Siri whenever system state modifications (such as setting reminders) are needed.
Inference Providing via Model Manager Services
Beyond conversational delegation, Apple implemented an Inference Providing protocol inside Model Manager Services. This architectural pattern allows third-party endpoints to replace Apple's internal server-side reasoning engine:
- Planner Interception: The third-party model (demonstrated with GPT-5.6 Terra) receives Apple's native planner prompts.
- Tool Call Definitions: System tools and capability schemas are exposed to the external model.
- Privileged Actions: The model requests cross-app executions (e.g., querying Mail, parsing action points, sending Messages via Contacts).
- Local Presentation: Siri renders the completed action and synthesizes speech using native OS UI components.
Regulatory Context and Availability
Driven in part by European Union Digital Markets Act requirements for system-level interoperability, these protocols confirm Apple is standardizing tool-calling contracts across desktop and mobile. While user-facing switches in macOS Golden Gate RC currently expose only ChatGPT under 'Ask...' and Apple has not yet opened the model delegation entitlement to third parties, the compiled frameworks prove comprehensive cross-model agent support is already present in OS builds.
✓ When to use
- When evaluating future multi-agent operating system architectures and EU DMA compliance strategies.
✕ When NOT to use
- When designing production integrations expecting public API stability from Apple today.
What to do today
- Monitor Apple developer documentation for future public releases of Model Delegation entitlements.
Sources