Agent Development Kit (ADK) for Kotlin

An open-source, code-first Kotlin toolkit for building, evaluating, and deploying sophisticated AI agents with flexibility and control.

Important Links: Docs & Samples & Python ADK & Java ADK.

Agent Development Kit (ADK) is designed for developers seeking fine-grained control and flexibility when building advanced AI agents that are tightly integrated with services in Google Cloud. It allows you to define agent behavior, orchestration, and tool use directly in code, enabling robust debugging, versioning, and deployment anywhere – from your laptop to the cloud.

ADK for Kotlin now follows a stable release cycle, with regular releases published to Maven Central under semantic versioning.

✨ Key Features

• Rich Tool Ecosystem: Utilize pre-built tools, custom functions, OpenAPI specs, or integrate existing tools to give agents diverse capabilities, all for tight integration with the Google ecosystem.

• Code-First Development: Define agent logic, tools, and orchestration directly in Kotlin for ultimate flexibility, testability, and versioning.

• Modular Multi-Agent Systems: Design scalable applications by composing multiple specialized agents into flexible hierarchies.

• On-device & Cloud Agents on Android: Run agents fully on-device with LiteRT-LM (with tool calling) or Gemini Nano via ML Kit, reach the cloud with Firebase AI, and compose on-device and cloud models into a single hybrid system. See On-device and Cloud Agents on Android.

• First-Class Java Interop: Build the same agents from plain Java, not just Kotlin, with a familiar, idiomatic API and no coroutines to learn. See Using ADK from Java.

🚀 Installation

If you're using Maven, add the following to your dependencies:

<dependency>
  <groupId>com.google.adk</groupId>
  <artifactId>google-adk-kotlin-core-jvm</artifactId>
  <version>1.4.0</version>
</dependency>

If you're using Gradle:

implementation("com.google.adk:google-adk-kotlin-core:1.4.0")

📦 Modules

Every module is published under the com.google.adk group and shares the version shown above.

📚 Documentation

For building, evaluating, and deploying agents by follow the Kotlin documentation & samples:

• Documentation
• Samples

🏁 Feature Highlight

Same Features & Familiar Interface As Python ADK:

import com.google.adk.kt.agents.LlmAgent
import com.google.adk.kt.agents.Instruction
import com.google.adk.kt.models.Gemini
import com.google.adk.kt.tools.GoogleSearchTool

val rootAgent = LlmAgent(
    name = "search_assistant",
    description = "An assistant that can search the web.",
    model = Gemini(name = "gemini-3.1-flash-lite-preview"),
    instruction = Instruction("You are a helpful assistant. Answer user questions using Google Search when needed."),
    tools = listOf(GoogleSearchTool())
)

GenAI SDK based Gemini currently prevents usage of API_KEY and GoogleCredentials on Android. Use Firebase AI instead.

☕ Using ADK from Java

ADK Kotlin is built to be called from plain Java too, not only from Kotlin. Every builder-style API mirrors the Kotlin one, so the search agent above is the same agent in Java:

import com.google.adk.kt.agents.BaseAgent;
import com.google.adk.kt.agents.LlmAgent;
import com.google.adk.kt.models.Gemini;
import com.google.adk.kt.tools.GoogleSearchTool;

BaseAgent rootAgent =
    LlmAgent.builder()
        .name("search_assistant")
        .description("An assistant that can search the web.")
        .model(new Gemini("gemini-3.1-flash-lite-preview"))
        .instruction(
            "You are a helpful assistant. Answer user questions using Google Search when needed.")
        .tools(GoogleSearchTool.builder().build())
        .build();

Coroutines stay out of the way. The com.google.adk.kt.interop package exposes Java-facing base types that replace the engine's suspend functions and Flow with CompletableFuture and Reactive Streams Publisher:

• PublisherRunner wraps a runner and streams events as a Publisher<Event>, which adapts in one line to RxJava (Flowable.fromPublisher(...)), Reactor (Flux.from(...)), or Spring WebFlux.
• CompletableFutureSessionService, CompletableFutureArtifactService, and CompletableFutureMemoryService wrap a session, artifact, or memory service so that each call returns a CompletableFuture, which Reactor adapts with Mono.fromFuture(...). Closing a wrapper also closes the service it wraps, if that service is AutoCloseable.
• BaseFutureTool, BaseFutureToolset, BaseFuturePlugin, BasePublisherModel, and the BaseFuture* session, memory, and artifact services let you implement ADK extension points in Java by overriding CompletableFuture- or Publisher-returning methods.
• BaseFutureLiveConnection (experimental) lets a Java model support live connections: subclass it and return an instance in a CompletableFuture from the model's BasePublisherModel.connectAsync override. Its sends and close return CompletableFutures, and each model turn is a Publisher.
• @Tool / @Param annotations work from Java: ReflectiveTools.fromMethod(...) turns an annotated plain method into a tool for javac-only modules, or KSP can process @Tool directly when you build with the Kotlin toolchain.

See examples/java/ for runnable ports of this repository's agents written entirely in Java.

📱 On-device and Cloud Agents on Android

ADK Kotlin agents run on Android either fully on-device — no API key and no network at inference time, for offline, low-latency, or privacy-sensitive use — or against the cloud. Every backend is a Model behind the same LlmAgent API, so switching between them is a one-line change.

• LiteRT-LM runs open models such as Gemma on-device through LiteRT-LM, Google's on-device inference framework, with full tool / function calling so an on-device agent can drive custom tools. It runs on both Android and the JVM (desktop). See litertlm/README.md.
• ML Kit (Gemini Nano) runs the built-in Gemini Nano model through the ML Kit GenAI Prompt API. It is Android-only and published as a -beta pre-release; tool calling is not supported yet (functionCall / functionResponse parts are dropped), so use it for plain chat/generation for now.
• Firebase AI reaches a cloud Gemini model through Firebase AI Logic — the recommended way to call Gemini from Android without embedding an API key — with full tool calling.

Hybrid on-device + cloud systems

Because on-device and cloud models are interchangeable Model implementations, you can compose them into a single multi-agent system: an on-device agent can handle offline, cheap, or privacy-sensitive turns and delegate harder tasks to a cloud agent, all within one LlmAgent hierarchy. See Modular Multi-Agent Systems for how agents are composed.

All three backends have runnable multi-turn chat examples in the examples/android/ Compose app — LiteRT-LM chat, ML Kit chat, and Firebase AI. See the Android examples README for how to build and run the app, obtain the on-device models, and configure Firebase.

Serving agents over HTTP

Add com.google.adk:google-adk-kotlin-webserver alongside the core artifact. AdkApiServer then serves the agent runtime contract — app discovery, sessions, artifacts, the run endpoints, health and version — without the development surface, so it can be deployed headlessly:

import com.google.adk.kt.webserver.AdkApiServer
import com.google.adk.kt.webserver.AdkServerConfig

fun main() = AdkApiServer(AdkServerConfig.inMemory(rootAgent)).start(wait = true)

That listens on port 8080. inMemory keeps session and artifact state in the process, so it suits a local run; build AdkServerConfig directly to serve several agents or to persist state.

Run it however you run any other main class. With the Gradle application plugin that is mainClass = "com.example.MainKt" — a top-level main in Main.kt compiles to MainKt, not Main — and then ./gradlew run.

The server binds loopback, because these endpoints are unauthenticated. Put your own authentication in front of it before widening that:

import com.google.adk.kt.webserver.AdkApiServer
import com.google.adk.kt.webserver.AdkServerConfig

fun main() =
  AdkApiServer(AdkServerConfig.inMemory(rootAgent).copy(host = "0.0.0.0")).start(wait = true)

Development UI

Same as the beloved Python Development UI. A built-in development UI to help you test, evaluate, debug, and showcase your agent(s).

AdkDevServer takes the same config and adds the UI at http://localhost:8080/dev-ui, together with the trace and agent-graph endpoints it drives:

import com.google.adk.kt.webserver.AdkServerConfig
import com.google.adk.kt.webserver.dev.AdkDevServer

fun main() = AdkDevServer(AdkServerConfig.inMemory(rootAgent)).start(wait = true)

To leave the UI unmounted, set the webUiEnabled field on AdkServerConfig:

import com.google.adk.kt.webserver.AdkServerConfig
import com.google.adk.kt.webserver.dev.AdkDevServer

fun main() =
  AdkDevServer(AdkServerConfig.inMemory(rootAgent).copy(webUiEnabled = false)).start(wait = true)

Three things decide whether the UI is mounted. Highest wins:

1. the adk.web.ui.enabled system property, when it is set to true or false in any casing — any other value is ignored with a warning
2. the webUiEnabled field on AdkServerConfig, when it is not null
3. the server class: AdkApiServer leaves the UI unmounted, AdkDevServer mounts it

So the system property is not an alternative to the field, it outranks it: a stray -Dadk.web.ui.enabled=true in the launch environment re-mounts the UI even though the config says webUiEnabled = false. It ranks highest so that a deployment which cannot change code can still turn the UI off.

To try a few of this repository's own agents rather than your own:

GOOGLE_API_KEY=... ./gradlew :google-adk-kotlin-examples:runServer --args="--dev"

📂 Examples

The agent snippet above is the short version. Every runnable example lives under the examples directory of this repository.

• examples/ — JVM examples.

• examples/java/ — the same agents written in plain Java, compiled against the published JVM artifacts to exercise the Java-facing surface.

• examples/android/ — a Compose app showing the Android side, including the on-device (LiteRT-LM, ML Kit) and cloud (Firebase AI) agents described in On-device and Cloud Agents on Android.

🤝 Contributing

We welcome contributions from the community! Whether it's bug reports, feature requests, documentation improvements, or code contributions, please see our CONTRIBUTING.md to get started.

📄 License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.

Preview

This feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms. Pre-GA features are available "as is" and might have limited support. For more information, see the launch stage descriptions.

Happy Agent Building!