Android exposes AppFunctions to OS agents

Google’s Android developer documentation says AppFunctions is now the mechanism for Android apps to expose actions, data and services to OS-level agents and assistants, including Gemini, on devices running Android 16 or higher. Google describes the feature as an Android platform API with a Jetpack library that lets apps behave like on-device Model Context Protocol, or MCP, servers. The company says authorized callers can discover and execute those functions through natural-language requests, while execution happens locally on the device. Google says AppFunctions integration with Gemini is in private preview with trusted testers, though developers can begin building against the APIs now. ### What exactly did Android open up to agents? Android’s AppFunctions overview says developers can define self-describing functions that expose an app’s capabilities as orchestratable tools for “agents and assistants, like Google Gemini.” Those functions are registered with the Android operating system, which then allows qualified callers to invoke them for user tasks such as creating reminders, assembling playlists or retrieving app content. (developer.android.com) Google’s February 25, 2026 Android Developers Blog post said the shift is from users opening apps step by step to asking AI systems to complete tasks for them. In that post, Matthew McCullough, Google’s vice president of product management for Android development, said Android was introducing “early stage developer capabilities” to connect apps with agentic apps and personalized assistants. (developer.android.com) ### How is this different from cloud-based AI tool calling? Google’s documentation says AppFunctions is the “mobile equivalent” of tools within MCP, but implemented on-device rather than through server-side tool endpoints. The Android Developers Blog said AppFunctions mirrors backend MCP servers while providing “an on-device solution for Android apps,” and compared it with WebMCP while noting that the functions execute locally on the device. (android-developers.googleblog.com) Android’s AI landing page says the broader platform strategy combines AppFunctions with on-device models such as Gemini Nano, ML Kit GenAI APIs and, where needed, cloud or hybrid AI through Firebase AI Logic. The same page says UI automation can serve as a fallback on supported devices when an app has not added AppFunctions. ### What does Google say about privacy, permissions and speed? Google’s February 2026 blog post said the company was designing these capabilities “with privacy and security at their core” while it remains in early beta stages. (developer.android.com) The AppFunctions overview says only callers with the `EXECUTE_APP_FUNCTIONS` permission can discover and execute AppFunctions, which limits access to authorized apps, agents and assistants. (developer.android.com) Android’s public docs do not state a millisecond performance target, but they repeatedly say execution is local to the device and tied to Android’s system registry for app functions. That on-device design is the basis for developer commentary this week describing faster reactions and less need to send requests to a server, though Google’s own documentation stops short of quantifying latency. ### Where is Google already showing this in use? (android-developers.googleblog.com) Samsung’s Galaxy S26 series was the example Google highlighted in its February blog post. Google said Samsung Gallery uses AppFunctions with Gemini so a user can ask for photos — such as pictures of a cat — and Gemini can identify the right function, call Samsung Gallery and return the results inside the Gemini app without forcing the user to switch apps. (developer.android.com) Google also said in that post that Gemini can already automate tasks across categories including Calendar, Notes and Tasks on devices from multiple manufacturers. The company added that the Samsung Gallery integration would expand to Samsung devices running OneUI 8.5 and higher. ### What do developers need to ship this? Google’s setup guide says AppFunctions requires apps to compile with API level 36 or higher, while runtime support is handled by the Jetpack library. (android-developers.googleblog.com) The guide lists `androidx.appfunctions:appfunctions`, `appfunctions-service`, and the `appfunctions-compiler` artifact for Kotlin Symbol Processing, with the current documentation showing version `1.0.0-alpha09`. Android’s docs also point developers to analysis and code-generation tools that scan an app for candidate workflows, generate Kotlin code for AppFunctions, and refine KDoc so agents can understand each function’s purpose and parameters. Google says those tools are meant to help developers identify high-value actions such as search, create, share and task completion flows that can be invoked without opening the full app interface. (developer.android.com) ### What happens next for Android developers? As of May 2026, Google says AppFunctions with Gemini remains in private preview with trusted testers, while the public developer documentation, setup guide and API references are already live. Google’s AI on Android page directs developers to prepare apps now using AppFunctions, and the Android Developers Blog says the company is also building a separate UI automation framework for OEM-preloaded assistants on supported devices. (developer.android.com 1) (developer.android.com 2)