Kotlin Multiplatform port of Vercel AI SDK v6 patterns and public feature areas.

This is a green-room Kotlin rewrite of the architectural contracts, not a TypeScript source translation and not an official Vercel package.

• Common Kotlin API for shared agent code.
• Android library publication.
• iOS Kotlin Multiplatform klibs for iosX64, iosArm64, and iosSimulatorArm64 published through Maven; the XCFramework is CI-built but not distributed.
• JVM artifact for backend and desktop services.
• Linux/Native linuxX64 artifact for server-side Kotlin/Native and CLI consumers.

Use the published beta from Maven Central:

dependencies {
    implementation("ai.torad:torad-aisdk:0.3.0-beta01")
}

For local development against a checkout, use a source dependency or publishToMavenLocal:

./gradlew publishToMavenLocal

import ai.torad.aisdk.LanguageModel
import ai.torad.aisdk.StepCountIs
import ai.torad.aisdk.Tool
import ai.torad.aisdk.ToolLoopAgent
import ai.torad.aisdk.ToolSet
import ai.torad.aisdk.providers.MockLanguageModelTextOnly
import kotlinx.coroutines.flow.first
import kotlinx.serialization.Serializable
import kotlinx.serialization.serializer

@Serializable
data class EmptyInput(val unused: String = "")

val helloTool = Tool<EmptyInput, String, Unit>(
    name = "hello",
    description = "Return a greeting.",
    inputSerializer = serializer(),
    outputSerializer = serializer(),
) { _ -> "Hello from a tool." }

class HelloAgent(model: LanguageModel, tools: ToolSet<Unit>) :
    ToolLoopAgent<Unit, String>(
        model = model,
        instructions = "Be brief.",
        tools = tools,
        stopWhen = StepCountIs(3),
    )

suspend fun main() {
    val agent = HelloAgent(
        model = MockLanguageModelTextOnly("Welcome."),
        tools = ToolSet(helloTool),
    )

    val result = agent.generate(prompt = "Say hi").first()
    println(result.text)
}

Send application prompts through Agent.generate / Agent.stream, or use TextGenerator(model).generate(...), TextGenerator(model).stream(...), and TextGenerator(model).streamResult(...) for direct text calls. Direct LanguageModel.generate, LanguageModel.stream, and LanguageModel.streamResult calls are still supported for provider authors and low-level integrations, but they require @OptIn(LowLevelLanguageModelApi::class).

• Agent and ToolLoopAgent.
• Typed Tool() definitions, DynamicTool(), schemas, and ToolSet.
• TextGenerator(model).generate(...), TextGenerator(model).stream(...), and cold Flow<StreamEvent> streaming.
• Structured output through Output.obj, Output.array, Output.choice, and Output.json.
• Structured text helpers through TextGenerator.streamResult(...) and StreamTextResult.
• Embeddings, reranking, image generation, speech generation, transcription, and video generation model contracts.
• Provider registry and Provider(...) routing.
• Gateway facade with Gateway(), gateway, gateway metadata APIs, gateway errors, provider-executed gateway tool descriptors, and a Ktor-backed KtorGatewayTransport.
• OpenAI-compatible Ktor provider for chat, completions, embeddings, images, speech, and transcription through OpenAICompatible(...).
• LiteRT-LM adapter for on-device Android/JVM inference via LiteRTLanguageModel, preserving SDK-owned generate, stream, reasoning-channel, and tool-loop routing.
• Provider-utils parity helpers: schemas, IDs, JSON event stream parsing, headers, base64 byte helpers, media and URL validation utilities.
• Text stream, UI message stream, and chat transport primitives for Kotlin hosts.
• Telemetry helpers, global/local telemetry integrations, and a KMP tracer/span abstraction.
• Compatibility helpers such as DefaultGeneratedFile, pruneMessages, experimental media aliases, and v6 public error types.
• Lifecycle hooks, middleware, stop conditions, call/step preparation, and cancellation.
• UI message aggregation types for Compose, SwiftUI, or server-rendered hosts.
• Mock models for deterministic tests across every model family.

Provider facades for the AI SDK v6 package ecosystem are folded into this root artifact for now. Future aisdk-provider-* artifacts can split publication boundaries without changing the common contracts.

• Docs wiki
• Local LLM context
• Interface contract
• Evolution policy
• Port notes
• Architecture decisions
• Parity ledgers
• Kotlin SDK engineering standard

Use JDK 21 to build the project. JVM and Android bytecode target JVM 17. Run tools/bootstrap once in a fresh clone to activate the committed pre-commit gates.

./gradlew jvmTest
./gradlew check publishToMavenLocal

Android publication requires an Android SDK with compile SDK 36 installed.

Apache-2.0. See LICENSE.