flutter_agent_kit 0.0.1

Flutter bindings for Google's Agent Development Kit (ADK) for Kotlin — build on-device and cloud AI agents with tool calling, from Dart.

Flutter bindings for Google's Agent Development Kit (ADK) for Kotlin — build on-device and cloud AI agents with tool calling, from pure Dart.

Status: Android-only v1. Not yet published to pub.dev.

Features

• Streaming chat, token by token, against a real model — not a mock.
• Tools defined entirely in Dart — a name, a JSON-Schema parameter map, and a callback. No native code required.
• Two model backends: AgentModelConfig.cloud() (Firebase AI-backed Gemini) or AgentModelConfig.onDevice() (LiteRT-LM, offline). Both support tool calling identically, so switching backends never silently breaks a tool.
• Session persistence as a single config flag (AgentSessionConfig.memory() vs .persistent()) — no separate session-service concept to learn.

See docs/adk-kotlin-capability-map.md for exactly what ADK-Kotlin can and can't do, and why this package is shaped the way it is.

Getting started

Add the dependency:

dependencies:
  flutter_agent_kit: ^0.0.1

If you're using AgentModelConfig.cloud(), your app needs its own Firebase project (this package can't supply one for you):

1. Create a project at console.firebase.google.com and register your Android app.
2. Download google-services.json into android/app/.
3. In the project's console, go to AI Services → AI Logic, click Get started, and choose Gemini Developer API (free tier).

Android requires compileSdk 37, minSdk 26, and a jvmToolchain(17) — these come from ADK-Kotlin's and Firebase's own floors, not this package. See flutter_agent_kit_android's README for the full setup and troubleshooting notes.

Usage

import 'dart:math';
import 'package:flutter_agent_kit/flutter_agent_kit.dart';

final agent = Agent(
  instructions: 'You are a helpful assistant. Use tools when they help.',
  model: const AgentModelConfig.cloud(),
  tools: [
    AgentTool(
      name: 'roll_dice',
      description: 'Rolls a standard six-sided die and returns the result.',
      parametersSchema: const {'type': 'object', 'properties': {}},
      execute: (args) async => {'result': Random().nextInt(6) + 1},
    ),
  ],
);

await for (final event in agent.send('Roll a die for me')) {
  switch (event) {
    case AgentTextDelta(:final text):
      print('...$text'); // partial, streamed chunk
    case AgentFinalResponse(:final text):
      print('Final: $text');
    case AgentToolCall(:final toolName):
      print('Calling $toolName...');
    case AgentError(:final message):
      print('Error: $message');
  }
}

await agent.dispose(); // when the agent is no longer needed

To remember the conversation across app restarts, pass session: const AgentSessionConfig.persistent() — everything else about agent.send(...) stays the same.

Additional information

• flutter_agent_kit/example is a full working chat app (streaming, the tool round-trip, a persistence toggle) — the fastest way to see this package actually run.
• docs/ at the repo root has the full research trail: what ADK-Kotlin supports, how it compares to Apple's on-device framework, and the implementation plan this package was built against.
• flutter_agent_kit_platform_interface is the platform contract this package talks to — only relevant if you're implementing support for a new platform.