AIツール通信
ニュース📖 16分で読める

Siri AI Is Open to All Developers: 40 Tools, 32 Calls, and Hard Limits

Apple opened Siri AI to every app via App Intents in the iOS 27.3 betas. How ToolIntent, SKILL.md app skills, the 40-tool cap and spring 2027 submission window actually work.

#Siri AI#App Intents#iOS 27.3
Siri AI Is Open to All Developers: 40 Tools, 32 Calls, and Hard Limits

The short version

On November 3, 2026, Apple published "Siri AI is now open to all developers" and shipped a Siri AI Developer Preview inside the iOS 27.3, iPadOS 27.3, macOS 27.3, and visionOS 27.3 betas. Any app can now expose its content and actions to Siri through App Intents — not just apps that fit one of Apple's predefined schemas.

Read the fine print before you plan a launch around it. This is a developer preview, not a feature release.

  • Integrations run only in development-signed builds and internal TestFlight. External TestFlight and App Store builds do not get them.
  • Customers see this "in a few months." App Store submissions open in spring 2027.
  • Terms and usage policies for Siri AI experiences have not been published.
  • The App Intents API and Siri's behavior may change during the preview, per Apple's own documentation.

For anyone building right now, the work is: decide which actions your app should hand to Siri, model them cleanly, and measure them before the spring window opens.

What actually changed

Until this preview, Siri AI's natural-language reach was limited to a small set of Apple-defined domains — calendar, reminders, messages, and similar. If your app did not fit a schema or implement an App Shortcut, Siri could not touch it.

Four things changed:

  • Custom actions become tools. Add the @AppIntent(.tool) Swift macro to an existing app intent and it conforms to ToolIntent, making it callable by Siri. No new entitlement, no separate developer program. Intents that already conform to a Siri-supported schema should not get the macro — that is a build error.
  • App skills. Markdown instructions that teach Siri how to combine several actions into one experience. Apple describes this as a subset of the Agent Skills format — the same SKILL.md shape Claude Code and Codex use.
  • The Spotlight semantic index opens up. Siri can now surface any entity an app indexes, not only entities that fit an existing schema.
  • The app name becomes optional in three situations: when the app is in the foreground, when its Live Activity is running, or when the user is already in a Siri conversation about that app. Background apps still need to be named.
  • Apple's example is the useful one: request a ride, then say "Hey Siri, change my ride pickup to a new address" while the ride's Live Activity is on screen. The ride-hailing app never gets opened.

    Apple's own apps got the same treatment across 15 first-party apps — Calendar, Files, Finder, Journal, Mail, Maps, Messages, Music, Notes, Photos, Preview, Reminders, Safari, Shortcuts, and Weather. 9to5Mac's early testing found Siri AI can now schedule and edit iMessages, build and save Music playlists, create Calendar events with invitees and video links, rename people and pets in Photos, and sort or close Safari tabs.

    The trap: skills are guidance, not a workflow engine

    This is the part that will burn teams who treat skills as a replacement for a deterministic pipeline.

    An app skill is a SKILL.md inside a Skills.skillsbundle (exactly one bundle per app target). The frontmatter supports only name and description. The body is capped at 10,000 characters. No scripts, no reference files, no other files in the directory.

    Apple states the constraint plainly:

    You can't use app skills to enforce deterministic behavior.

    And:

    Siri AI doesn't sequentially perform the tool intents that your skill lists.

    Siri reads the metadata, decides whether the skill matches the request, and only then reads the body — as guidance for planning, not as a script to execute. If a step must happen, enforce it in code.

    Two numbers worth knowing, and worth treating as provisional: 40 tools per app and 32 tool calls chained per skill. Both came from Apple's Apple Intelligence Group Labs sessions on November 4–5 and were reported by MacStories' John Voorhees on November 6. Neither figure appears in the published documentation as of this writing, which means they can move.

    Design your intents for a planner, not a menu

    The single highest-leverage decision is granularity. A planner can only call what it can see, so a single perform() that checks status, changes an address, notifies the user, and charges a card is invisible in its middle steps. Split it:

    • "Check the downtown delivery" is one tool. "Move the pickup to 500 Market" is another.
    • Use an AppEntity parameter when the value already exists in your data and an entity query or the semantic index can find it.
    • Use String, Date, or a small app enum when the person is dictating a new value.
    • Consider @AppEntity(.generable) on a TransientAppEntity when the model can fill every property straight from the request without a lookup — Apple's flight-search example is the canonical case.

    Descriptions are load-bearing. Siri relies on the title and description of the intent, of each parameter, and of each @Property/@DeferredProperty/@ComputedProperty on your entities to disambiguate. An intent with a missing description is an intent Siri will guess at.

    Verification is now a three-stage pipeline

    Apple added two layers between your unit tests and manual testing:

    • Siri Integration Advisor in Xcode 27.3 reviews your App Intents setup and suggests improvements.
    • SiriEvaluations lets you script multi-turn requests, then measure which intents were called, which values were passed, which entities were found, which skills loaded, and what Siri said.

    The Loads check confirms a skill was actually loaded into planning for a turn — and only a required load counts toward the skillsLoadedPass metric, because a preferred load might be satisfied by calling tools directly. Pair Loads with Calls for the tools you expect, or a passing skill load tells you nothing about whether the work happened.

    Because responses vary run to run, the right practice is to run the same scenario several times and look for stable passes, not a single green run.

    Sensitive data: the app entity defaults are open

    App entities do not require authentication by default before Shortcuts or Siri present their information. For anything financial or personal, set an IntentAuthenticationPolicy — .requiresAuthentication or .requiresLocalDeviceAuthentication — either on the entity as a whole or per property.

    Two caveats that matter:

    • Property-level policies govern Siri's responses only. If an intent returns sensitive data, require authentication at the intent level too.
    • Set visibility: .internal on properties that should never leave your app and its extensions. Siri, Shortcuts, and other system features will not read them.
    • Donate intents to the system as users perform actions in your app; behavioral cues improve disambiguation of ambiguous requests.

    What it means for teams outside Cupertino

    Language coverage is real. Siri AI runs in English, French, Japanese, Korean, Portuguese, and Spanish, with Traditional Chinese, Norwegian, Turkish, and Vietnamese announced. Japanese beta support arrived with iOS 27.2 on October 26, so Japanese-speaking developers can test custom app actions today. The EU is excluded. Siri AI is currently unavailable in the EU on iOS, iPadOS, and watchOS. If you ship in Europe, you cannot measure adoption with local users before spring. Plan for a two-speed rollout. A fee is coming. Apple's own footnote says server-side Apple Intelligence features — Siri AI among them — are subject to daily usage limits, and that "expanded access to such features will be available for a fee in the future." No price, no feature list. Tool requests run on Private Cloud Compute, so the per-request cost is Apple's, and the monetization question is not settled. Model that into your business case rather than assuming free unlimited agent traffic. Users under 13 are excluded.

    What to do this month

  • Audit your App Intents. List every intent you ship and check for missing title and description values. This is cheap and it directly determines whether Siri can find your actions.
  • Promote three to five actions, not forty. Ship the ones a person would ask for by name. You can widen later.
  • Write one skill and measure whether it helps. Compare a request handled by tools alone against the same request with a skill loaded, using SiriEvaluations.
  • Protect irreversible actions in code. Refunds, deletions, sends, and payments belong behind authentication and a confirmation surface, not behind a "do not do this" line in a Markdown file.
  • Read Apple's terms when they land before you commit to a customer-facing launch date.
  • Siri moved from "fixed phrases that open a shortcut" to "an agent that calls your app's functions from a natural-language request." That is a genuine platform shift. It is also, as of November 12, 2026, a preview that no user can touch. The teams that benefit are the ones that use the preview window to get their tool boundaries right — because when submissions open in spring 2027, the apps that get picked will be the ones whose actions Siri can actually find and call.

    🔒
    PR
    ABLENET レンタルサーバー
    国内データセンター。WordPress簡単インストール、24時間監視。
    この記事をシェア