简介

使用者痛点：开发者往往需要为每个 AI 助手单独实现一套接口，导致代码重复、维护成本高且难以跨网站。

谷歌刚刚发布了 AppFunctions 的首个公开文档和首个可用 alpha 版本。AppFunctions 是一个 Jetpack 库，它与一个新的 Android 网站 API 配合使用。可以将你应用的部分功能转化为 AI 助手可以发现和执行的工具。

使用者痛点：现有方案只能在第一方环境内工作。第三方应用必须为每个助手编写专属适配层，导致集成碎片化。

明确指出这里的新特性很有帮助，因为功能本身并不新…，真正的新特性是它的通用性。

之前，让助手在应用内代表你执行操作几乎完全是第一方的功能：你可以让 Google Assistant 在 Google 日历中添加事件。让 Bixby 在三星日历中创建事件，或者让 Siri 在苹果日历中执行相同的操作。每个助手都与其各自厂商的应用紧密集成。

第三方应用并非完全被拒之门外但每扇门的开启方式都各不相同。说起来，Android 的 App Actions 允许你将一组固定的内置意图暴露给 Google Assistant；苹果的 App Intents / SiriKit 为 Siri 提供了差不多功能；Bixby 则有其 Capsules。

每个助手都针对特定应用。通常仅限于预定义的词汇表，而且无法跨网站移植。真正意义上的助手控制要么是厂商的特权。要么是需要为每个助手编写很多集成代码，你的应用只能运行针对特定助手的代码。

这篇文章提供什么？按理说，

• 包含初始代码的仓库分支：-Starter-Code
• 包含 AppFunctions 功能的仓库分支：-Finished-Code
• 此处提供将 AppFunctions 添加到现有项目的 commit 提交

If you haven't used Model Context Protocol 。a quick primer will make everything that follows easier.

• MCP 服务器：a tiny process exposing a public list of tools. Each tool has a name,a plain‑language description,and a JSON schema for its parameters & return values.
• MCP 客户端： connects to server,asks “what tools do you have?”,reads descriptions,and decides which tool to call based on user's request.
• The dialogue is JSON‑RPC,carried over stdio or HTTP .

User pain point: Traditional MCP servers run **outside** your app – eir as a separate process or in cloud – so y lack privileged access to your app's internal state. Exposing a local MCP‑style tool often forces you to create an external API just for sake of agent.

• You annotate Kotlin methods with @AppFunction.
• The Jetpack annotation processor generates metadata that OS indexes at install time.
• A system service can discover those functions and invoke m on behalf of any caller that holds EXECUTE_APP_FUNCTIONS permission.

User Pain Points Resolved by These Three Properties:

1. No network round‑trip: The function runs locally on device,eliminating latency and server maintenance.
2. The OS indexes it automatically: You don’t need runtime registration code – just compile and ship.
3. KDoc becomes part of contract: When you set @AppFunction,your documentation is encoded into metadata and read by agents at runtime.

Minimum SDK requirements: compileSdk = X+,tagetSdk = X+,running on Android X+ devices.

秒内完成示例应用程序

The sample is a small Compose app that fetches jokes via Ktor,caches m with Room,and lets users favorite jokes. The core architecture is completely ordinary.

• Pain point: Third‑party apps traditionally expose functionality only through UI actions. Adding an AI‑visible entry required a completely new code path.
• This example adds one tiny function: **Clear all favorite jokes**. • User can tap a menu item. • An AI agent can call same function via @AppFunction.

The principle: **AppFunction is just a thin wrapper around existing business logic – not a duplicate implementation**.

第 1 步 — 添加依赖

appfunctions = ".-alpha09"
androidx-appfunctions = { module = "androidx.appfunctions:appfunctions"。version.ref = "appfunctions" }
androidx-appfunctions-service = { module = "androidx.appfunctions:appfunctions-service",version.ref = "appfunctions" }
androidx-appfunctions-compiler = { module = "androidx.appfunctions:appfunctions-compiler",version.ref = "appfunctions" }

// app/build.gradle.kts
ksp {
arg
}
dependencies {
implementation
implementation
ksp
}

• Pain point: forgetting to add KSP compiler means no metadata gets generated → functions invisible to agents.
• Pain point: using an old compileSdkVersion prevents library from compiling at all.

第 2 步 — 清单管道

• Pain point: omitting `` means OS never sees your metadata → no discoverability.
• Pain point: forgetting `BIND_APP_FUNCTION_SERVICE` permission blocks any system caller from binding to your service.

第 3 步 — 描述应用级元数据

This file lives in /res/xml/app_metadata.xml. It provides two key strings:

• Pain point : developers often write marketing copy here . That confuses LLMs because y look for actionable verbs rar than branding.
• Pain point : forgetting `displayDescription` results in users seeing nothing meaningful in Settings or Play Store deep links.

User‑focused tip:

• appfn:description : feed LLMs – keep it short,action‑oriented . ​ ​ ​ ​ ​ ​ ​ ​​ ​​ ​​

​​​

​​

---