Apple needs app tool APIs

Apple’s AI stack is getting more capable. But the connective tissue is still weirdly thin. Developers can now call Apple’s on-device model through the Foundation Models framework, and they can expose app actions through App Intents. What they still cannot do is register their app as a clean, reusable “tool” inside a general Apple Intelligence workflow. That missing layer is why people keep saying Apple needs app tool APIs. ### What’s the missing piece? A tool API would let an app describe what it can do in a structured way — fetch a task list, send money, book a table, edit a photo — and then let an AI system call that capability safely. Not by screen-scraping. Not by custom one-off deals. Just a standard contract. Apple has pieces of this today, but not the whole thing. App Intents exposes actions to Siri, Spotlight, widgets, Shortcuts, and other system surfaces. Foundation Models gives developers direct access to Apple’s on-device model. But there is no public layer that cleanly binds third-party app tools to model-driven planning. ### Doesn’t App Intents already do that? Sort of — but only sort of. App Intents lets apps publish actions and entities so Siri and Shortcuts can discover them. Apple explicitly says those intents can work with Siri’s action-taking features and with Apple Intelligence-enhanced system experiences. That is useful. But App Intents is still framed as system integration for actions, search, widgets, and automation. It is not presented as a general-purpose tool registry for any model-driven agent to call with standardized schemas, permission prompts, and failure handling. (developer.apple.com) ### What changed on Apple’s side? WWDC 2025 made the gap more obvious. Apple opened the Foundation Models framework so developers can tap the on-device model directly, and Apple’s developer pages now say features like guided generation and tool calling are built into that framework. That sounds like the platform is thinking in agent terms already. But the public docs stop short of saying third-party apps can plug their own tools into that tool-calling layer as first-class citizens. Developers can call the model, and developers can publish intents — but those are still two different roads. (developer.apple.com) ### Why is that separation a problem? Because an agent is only as useful as the actions it can reliably take. If the model can reason but has no standard way to invoke apps, every serious workflow turns into glue code. One partner gets a custom integration. Another app gets a Siri shortcut. A third gets some private handoff. That does not scale. The whole point of a platform API is to stop rebuilding the bridge for every app and every model. Apple’s own docs already hint at this split: apps can “tap into models directly” and also “integrate into places across the system via App Intents.” Those are complementary powers, but they are not yet one coherent tool layer. (developer.apple.com) ### Why does permissions design matter so much? Because tool use is where AI stops being a toy and starts touching real data. If a model can read your calendar, message a contact, or move money, the platform needs a standard way to ask for consent, scope access, explain what happened, and recover from errors. Apple is usually strongest when it turns messy capability into a predictable system contract. That is what made things like HealthKit, HomeKit, and App Intents legible to developers in the first place. (developer.apple.com) A tool API for Apple Intelligence would need the same shape — schemas, guardrails, and system-owned permission flows. That part is not public today. ### Could Apple be building toward this anyway? Probably. The ingredients are lining up. App Intents already gives Apple a structured action layer. Visual Intelligence uses App Intents to exchange information with apps. Shortcuts now has a “Use Model” action that can feed model output into broader workflows. And Foundation Models brings tool calling into the developer stack. Put those together and you can see the outline of a future system where apps register capabilities once and any approved Apple Intelligence surface can use them. (developer.apple.com) But that last step — the formal public contract — still looks missing. ### So what’s the real argument here? Basically, Apple does not just need better models. It needs better plumbing. If Apple Intelligence is going to become a real action layer across the iPhone, iPad, and Mac, third-party apps need a standard way to become callable tools. Without that, the platform stays fragmented — smart model here, app action there, lots of hand-built glue in between. With it, Apple could turn scattered AI features into an extensible ecosystem. That is the difference between demos and a platform. (developer.apple.com)