agents_genkit

Genkit adapters for the agents package.

Wire Genkit flows and tools into the full agent ecosystem — sessions, skills, context providers, DI — with zero changes to your existing agent code.

Features

Class	What it does
`GenkitFlowAgent`	Wraps a Genkit `Flow` (or any async function) as a first-class `AIAgent`
`GenkitToolsContextProvider`	Injects `AIFunction`s into every invocation via the `AIContextProvider` pipeline
`GenkitSkillsSource`	Exposes Genkit `Tool`s as `AgentSkill`s alongside file-based skill sources
`addGenkitAgent()`	One-call DI wiring: registers a `ChatClient`, function invocation, and a named `AIAgent`

Installation

dependencies:
  agents_genkit: ^0.1.0

Usage

Wrap a Genkit flow as an agent

import 'package:agents_genkit/agents_genkit.dart';
import 'package:genkit/genkit.dart';

final genkit = Genkit(plugins: [googleAI()]);

final summariseFlow = genkit.defineFlow(
  name: 'summarise',
  fn: (input, _) async {
    final response = await genkit.generate(
      model: googleAI.gemini('gemini-2.5-flash'),
      prompt: 'Summarise this in one sentence: $input',
    );
    return response.text;
  },
);

final agent = GenkitFlowAgent(
  name: 'summariser',
  description: 'Summarises text into one sentence.',
  run: (input) => summariseFlow(input),
);

final session = await agent.createSession();
final response = await agent.run(session, null, message: 'Dart is awesome...');
print(response.text);

Add Genkit tools to an existing agent

GenkitToolsContextProvider injects AIFunctions at invocation time so any ChatClientAgent gains them without modifying its static configuration.

import 'package:agents/agents.dart';
import 'package:agents_genkit/agents_genkit.dart';
import 'package:extensions/ai.dart';

// Define a function with the extensions AI abstraction.
final getWeather = AIFunction(
  name: 'get_weather',
  description: 'Returns the current weather for a city.',
  parameters: /* your schema */,
  fn: (args) async => fetchWeather(args['city'] as String),
);

// Inject it into any ChatClientAgent via HarnessAgentOptions.
final options = HarnessAgentOptions(
  aiContextProviders: [
    GenkitToolsContextProvider(functions: [getWeather]),
  ],
);

final agent = chatClient.asHarnessAgent(
  1_000_000,
  8_192,
  options: options,
);

Expose Genkit tools as skills

GenkitSkillsSource works with AggregatingAgentSkillsSource so Genkit tools appear alongside your existing file-based skills.

import 'package:agents/agents.dart';
import 'package:agents_genkit/agents_genkit.dart';
import 'package:genkit/genkit.dart';

final genkit = Genkit(plugins: [googleAI()]);

final calendarTool = genkit.defineTool(
  name: 'create_calendar_event',
  description: 'Creates a Google Calendar event.',
  fn: (input, _) async => createEvent(input),
);

final skillsSource = AggregatingAgentSkillsSource([
  AgentFileSkillsSource(directory: Directory('skills')),
  GenkitSkillsSource(tools: [calendarTool]),
]);

final options = HarnessAgentOptions(agentSkillsSource: skillsSource);

DI registration with addGenkitAgent()

For hosted applications using the extensions DI container, addGenkitAgent() mirrors the familiar addGemma4LiteRtAgent() pattern.

import 'package:agents_genkit/agents_genkit.dart';
import 'package:extensions/hosting.dart';
import 'package:genkit/genkit.dart';

void main() async {
  final builder = HostApplicationBuilder();

  builder.services
    .addGenkitAgent(
      model: googleAI.gemini('gemini-2.5-flash'),
      agentName: 'assistant',
      instructions: 'You are a helpful assistant.',
      genkit: Genkit(plugins: [googleAI()]),
    )
    .withInMemorySessionStore();

  await builder.build().run();
}

Additional information

Part of the agents monorepo.
File issues and feature requests on GitHub.
See the agents package for the full agent session, workflow, and middleware documentation.