Android 17 turns every app into an agent-callable tool

Google released Android 17 on June 17, shipping a platform API called AppFunctions that allows developers to expose their app’s native capabilities as orchestratable tools for on-device AI agents. The company described the release as the start of a transition from an operating system to what it calls an “intelligence system.”

The mechanics are straightforward. A developer annotates a Kotlin class with @AppFunction, adds KDoc comments describing parameters and behavior, and the Android platform registers those functions as discoverable tools. Android MCP, Google’s on-device implementation of the Model Context Protocol (MCP, Anthropic’s protocol for tool calling), then surfaces those functions to any agent or assistant running locally. Gemini is the named first consumer. An AI agent can call a note app’s createNote function, a calendar app’s scheduling function, or any other annotated capability without the developer building a separate integration per assistant.

This matters because it standardizes the discovery layer. Before this, every agent integration required custom work from each app developer: a plugin, a connector, or a custom API. AppFunctions collapses that into a single annotated class. The Jetpack library handling this is currently in alpha, and the Gemini integration is in private preview with a limited set of early-access testers. Google has not disclosed when the Gemini integration will reach general availability.

The OS-as-agent-platform angle is worth naming directly. Microsoft has been pushing a similar model through Windows with Copilot runtime integrations, and Apple’s App Intents framework already lets Siri and on-device models call app actions. Android MCP is Google’s structural answer to that: a protocol-level abstraction that any compliant agent can use, not just Gemini. If the protocol holds open, third-party agents could discover and call AppFunctions on Android without going through Google’s own assistant stack. The source blog post from the Android Developers Blog does not confirm this explicitly.

The comparison to Apple is the sharper one. Apple’s App Intents are tightly scoped to Apple Intelligence and Siri. Google’s framing around MCP as a named protocol standard signals a broader ambition: to make Android the platform where agent interoperability actually works across tools from different vendors. Whether that plays out depends on whether Google enforces openness or quietly gates the discovery channel to Gemini.

One skepticism note: the AppFunctions Jetpack library is alpha, and the agent skill Google ships to automate code generation for it is itself AI-generated boilerplate. The release announcement does not include independent data on function-calling accuracy, latency, or what happens when an agent calls a function incorrectly and the app state corrupts. These are the failure modes that matter for any production deployment.

The larger structural shift is real regardless. Android is now a platform where apps gain a second surface area: the traditional UI for humans, and an AppFunctions layer for agents. Developers who want their apps to be reachable inside agent workflows need to start building that layer now. The early access program for Gemini integration is at goo.gle/eap-af, which is the practical next step for any Android team evaluating this.

Teams building agentic products on Android should evaluate AppFunctions before committing to custom deep-link or API-based agent integrations; the protocol-native path will likely become the default discovery mechanism within two to three major Android releases.

Sourced from the Android Developers Blog (Google), published June 17, 2026.