Developing Genkit Dart

GoogleOfficialPublished by Google

Skill
Tool integration
Software development
Apache-2.0
Low risk

From the author

Generates code and provides documentation for the Genkit Dart SDK. Use when the user asks to build AI agents in Dart, use Genkit flows, or integrate LLMs into Dart/Flutter applications.

— the description by Google

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The Genkit Dart skill provides comprehensive guidance for utilizing the Genkit SDK to develop AI-powered applications. It covers the core framework, including defining flows and agents, managing sessions, handling persistence, and implementing generative UI (A2UI). The skill emphasizes the use of the Genkit CLI for local development, tracing, and debugging, while providing patterns for integrating various LLM plugins and managing type-safe schemas via the schemantic library.

What it does

Generates code and documentation for building AI agents and flows in Dart.
Provides patterns for integrating LLMs into Dart and Flutter applications.
Explains the use of the Genkit CLI for tracing, debugging, and running flows.
Details the implementation of persistent sessions, human-in-the-loop interrupts, and multi-agent orchestration.
Supports A2UI for streaming interactive UI surfaces to Flutter clients.
Guides the use of type-safe schemas with the schemantic library for tool and flow inputs.

When to use it

When building AI agents, chatbots, or assistants in Dart or Flutter.
When you need to integrate LLMs into a Dart backend or Flutter frontend.
When you need to debug model inputs, outputs, and tool calls using traces.
When implementing conversational workflows that require persistent state or human-in-the-loop approval.
When you want to stream interactive UI components from an agent to a Flutter app.

Advantages

Provides a unified interface for code generation, structured outputs, and agentic workflows.
Enables deep visibility into AI execution via the Genkit CLI and trace inspection.
Supports complex agentic behaviors like branching, background execution, and multi-agent delegation.
Facilitates type-safe development through integration with the schemantic library.
Offers a consistent developer experience across different LLM providers via a plugin ecosystem.

Usage examples

“How do I define a simple agent in Genkit Dart?”: The agent will provide a code snippet using ai.defineAgent, including system instructions, tools, and session storage configuration.
“How can I debug my Genkit flow locally?”: The agent will explain how to use 'genkit start -- dart run main.dart' to capture traces and inspect them via the Developer UI at localhost:4000.
“How do I use A2UI to stream a card to my Flutter app?”: The agent will guide you to register the A2uiPlugin, add the a2ui() middleware to your agent, and use the genui package on the Flutter client.
“How do I define a tool with a typed input schema?”: The agent will show how to use the @Schema() annotation and schemantic library to define a type-safe input class for your tool.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/google/developing-genkit-dart/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip developing-genkit-dart.zip -d .agents/skills

Documentation by Google

Genkit Dart

Genkit Dart is an AI SDK for Dart that provides a unified interface for code generation, structured outputs, tools, flows, and AI agents.

Core Features and Usage

If you need help with initializing Genkit (Genkit()), Generation (ai.generate), Tooling (ai.defineTool), Flows (ai.defineFlow), Embeddings (ai.embed), streaming, or calling remote flow endpoints, please load the core framework reference: references/genkit.md

1.0 vs pre-1.0 APIs

This skill targets Genkit Dart 1.0 (genkit: ^1.0.0). Older code and training data use pre-1.0 names that no longer compile; the common ones:

HTTP: startFlowServer / per-route shelfHandler / shelf_cors_headers are now GenkitRouter + serve() or asShelfHandler(), with CorsOptions; context providers take RequestData (genkit_shelf.md).
StatusCodes.UNAUTHENTICATED is StatusCode.unauthenticated (lowerCamelCase, same for EvalStatus.pass, ToolChoice.required).
No RetryPlugin; retry() is built in and takes Durations (initialDelay, maxDelay, also cacheTtl elsewhere).
Renames: GenerateResponseHelper -> GenerateResult (modelResponse, chunk.modelChunk), ExecutablePrompt -> Prompt<Input, Output>, embedMany -> embed, ToolFnArgs -> ToolFnArg, toolOutputSchema -> outputSchema, underlyingException -> cause, modelInfo: -> info:, RemoteAction.dispose() -> close(), prompt(input, PromptGenerateOptions(config: ...)) -> prompt(input, config: ...) (named options, like generate), defineMiddleware -> ai.defineGenerateMiddleware (app) / generateMiddleware (plugins).
Provider config types are prefixed: GeminiThinkingConfig, GeminiSafetySettings, GeminiGoogleSearch, GeminiSpeechConfig, GoogleGenAiEmbedderOptions, AnthropicThinkingConfig.
Removed: ToolInterruptException / ctx.interrupt() (return .interrupt(...)), GoogleAiModels / OpenAIModels constants (use googleAI.gemini('...'), openAI.model('...')). Simulated constrained output is opt-in via simulateConstrainedGeneration().
.prompt Picoschema is spec-compliant: name: type, description, never name(description): type (dotprompt.md).

Prompts (Dotprompt)

.prompt files keep prompt content out of Dart code with YAML frontmatter plus a Handlebars template. See references/dotprompt.md: promptDir, ai.prompt() (call/stream/render), variants, partials, named schemas via defineSchema, and the tools/maxTurns/returnToolRequests/use (middleware) frontmatter fields. A .prompt file can also back an agent directly via definePromptAgent.

Agents

Genkit Dart has an agent API for persistent, multi-turn conversations (sessions, snapshots, interrupts, branching, background execution, custom state, artifacts, and multi-agent delegation). The agent/session/snapshot APIs are experimental and live behind opt-in imports: server APIs come from package:genkit/experimental.dart (alongside package:genkit/genkit.dart), the browser/HTTP client from package:genkit/experimental_client.dart (alongside package:genkit/client.dart), and dart:io extras like FileSessionStore from package:genkit/experimental_io.dart. These entry points are @experimental, so importing them raises an experimental_member_use analyzer warning you can silence in analysis_options.yaml. The remoteAgent client works from any Dart app, including Flutter, and the backend is fully interchangeable — it can talk to a Genkit agent implemented in Dart, JS/TypeScript, or Go over the same HTTP protocol. A few Dart specifics: interrupts are ai.defineInterrupt(...) or tools that return .interrupt(...) (there is no ctx.interrupt()), sub-agent delegation uses the agents() middleware from package:genkit_middleware, and there is no artifacts() middleware yet (define artifact tools directly).

For more details see:

Agents: defining/serving an agent and client-managed state (start here).
Sessions & persistence: session stores (InMemorySessionStore/FileSessionStore/FirestoreSessionStore).
Human-in-the-loop / interrupts: pausing for approval/input via .interrupt(...) and resuming.
Branching: forking a conversation from a snapshot.
Background agents: detaching long-running turns and polling.
Working with state: typed custom session state, auto-synced to the client.
Artifacts: producing and reading named deliverables.
Multi-agent orchestration: delegating to sub-agents with the agents() middleware.
Advanced custom agents: defineCustomAgent for full turn control.
Deploying agents: serving agents over HTTP with GenkitRouter.addAgent (multiple agents, shelf, CORS).

Generative UI (A2UI)

Genkit Dart has an A2UI (Agent-to-UI) plugin (genkit_a2ui) that lets an agent stream interactive UI surfaces (cards, lists, forms, buttons), not just prose. The whole server-side integration is the a2ui() model middleware in an agent's (or ai.generate's) use list; the Flutter client renders surfaces with the genui package plus the helpers in package:genkit_a2ui/client.dart. Dart specific: you must register A2uiPlugin() in Genkit(plugins: [...]) (unlike JS, middleware is resolved by name from the registry).

A2UI: server middleware, options, Flutter/genui client rendering, user actions/forms, custom catalogs, and the security/trust boundary.

Genkit CLI (recommended)

genkit start unintrusively wraps any Dart program that uses the Genkit library, running it unchanged while capturing traces from every Genkit action so you can prove tools were actually called and inspect model I/O from the terminal, even for headless checks. It forwards stdio, so interactive CLI tools that rely on stdin/stdout work without issues. Running the app directly (dart run) skips trace capture, so you're debugging blind. Check install with genkit --version.

Installation:

curl -sL cli.genkit.dev | bash # Native CLI
# OR
npm install -g genkit-cli # Via npm
# OR run commands directly with npx without a global install (prefix every genkit command):
# npx genkit-cli start -- dart run main.dart

Primary pattern (default): prefix genkit start -- to your normal run command. This collects telemetry from any Genkit code your program runs, whether triggered from the dev UI, your own web server/web UI, or a plain script. Starts the Developer UI (usually http://localhost:4000) for running flows, model and agent playground, and browsing traces:

genkit start -- dart run main.dart
genkit start --noui -- dart run main.dart   # same, without the Dev UI (still a persistent server)

genkit start runs until you stop it with Ctrl+C. That is expected and correct for the common cases: a server your web/mobile app calls, or an interactive CLI you exit yourself. --noui only drops the Dev UI; it is not a one-shot command and will not exit on its own. Do not use genkit start as a blocking step in automated/non-interactive contexts; use flow:run (below) for that.

Non-interactive use (agents/CI): add the global --non-interactive flag before -- so the CLI uses defaults and never blocks on a prompt (e.g. the first-run analytics notice): genkit start --non-interactive -- dart run main.dart (works with flow:run too).

Run a flow (flow:run): invoke a specific flow by name from the CLI. Append your run command after -- to spin up the runtime just for this run (the command runs as-is to register your flows):

genkit flow:run myFlow '{"data": "input"}' -- dart run main.dart

This is self-terminating: it runs the flow once, prints a Trace ID, then exits, so it's the right choice for a quick, non-interactive check (unlike genkit start). Note: flow:run runs flows (ai.defineFlow), not agents; you can't flow:run an agent (ai.defineAgent) directly. To exercise an agent from the CLI, wrap one turn in a throwaway flow and run that (see Agents). Traces for this run can be inspected using the trace commands below.

Gotcha: top-level final declarations are lazy. Flows and agents defined as top-level final register with Genkit only when the symbol is first evaluated. An empty main() registers nothing, so flow:run fails with Process exited before runtime was ready. Reference the flow/agent symbols from main() (or import a module that does) so their define* calls actually run.

Debugging with traces: the fastest way to see prompts, model inputs/outputs, tool calls, latencies, and errors. Inspect from the terminal after any run under genkit start:

genkit trace:list                        # find recent trace IDs
genkit trace:get <traceId>               # full trace details (inputs, outputs, tool calls, errors)
genkit trace:get <traceId> --format json # machine-readable JSON, safe to pipe into jq or other parsers

For machine-readable output, pass --format json to get clean JSON you can pipe into jq or other parsers. The default output is human-oriented (banner/log lines, possible truncation on large traces), so don't pipe that form directly; use --format json, grep, or the Dev UI trace viewer.

Documentation:

genkit docs:search "streaming" dart
genkit docs:list dart
genkit docs:read dart/flows.md

Plugin Ecosystem

Genkit relies on a large suite of plugins to perform generative AI actions, interface with external LLMs, or host web servers.

When asked to use any given plugin, always verify usage by referring to its corresponding reference below. You should load the reference when you need to know the specific initialization arguments, tools, models, and usage patterns for the plugin:

Plugin Name	Reference Link	Description
`genkit_google_genai`	references/genkit_google_genai.md	Load for Google Gemini plugin interface usage.
`genkit_vertexai`	references/genkit_vertexai.md	Load for Gemini via Vertex AI (Google Cloud auth, versioned model IDs, embeddings incl. multimodal).
`genkit_anthropic`	references/genkit_anthropic.md	Load for Anthropic plugin interface for Claude models.
`genkit_openai`	references/genkit_openai.md	Load for OpenAI plugin interface for GPT models, Groq, and custom compatible endpoints.
`genkit_middleware`	references/genkit_middleware.md	Load for Tooling for specific agentic behavior: `filesystem`, `skills`, and `toolApproval` interrupts.
`genkit_mcp`	references/genkit_mcp.md	Load for Model Context Protocol integration (Server, Host, and Client capabilities).
`genkit_chrome`	references/genkit_chrome.md	Load for Running Gemini Nano locally inside the Chrome browser using the Prompt API.
`genkit_shelf`	references/genkit_shelf.md	Load for serving flows/models/agents over HTTP: `GenkitRouter` (core `package:genkit/io.dart`, standalone `dart:io`) and mounting it into Shelf apps.
`genkit_firebase_ai`	references/genkit_firebase_ai.md	Load for Firebase AI plugin interface (Gemini API via Vertex AI).
`genkit_a2ui`	references/a2ui.md	Load for A2UI (Agent-to-UI): streaming generative UI surfaces via the `a2ui()` middleware, rendered on the client with `genui`.

External Dependencies

Whenever you define schemas mapping inside of Tools, Flows, and Prompts, you must use the schemantic library. To learn how to use schemantic, ensure you read references/schemantic.md for how to implement type safe generated Dart code. This is particularly relevant when you encounter symbols like @Schema(), SchemanticType, or classes with the $ prefix. Genkit Dart uses schemantic for all of its data models so it's a CRITICAL skill to understand for using Genkit Dart.

Best Practices

Agent or flow? If the task is conversational, multi-turn, or described as "an agent", "assistant", or "chatbot", build it with ai.defineAgent (see Agents) rather than hand-rolling a generate + tools loop inside a flow. Reach for a plain flow only for single-shot, stateless generation.
Always check that code cleanly compiles using dart analyze before generating the final response.
Always use the Genkit CLI for local development and debugging.
Verify with traces, not a blind run. Running the app directly (dart run) does not capture dev traces. See the Genkit CLI section for how to run your app and capture traces.

Files

ATTRIBUTION.mdAdded by AutoGineer862 B
LICENSELicence file added by AutoGineer11.1 KiB
SKILL.md13.4 KiB
references/a2ui.md14.7 KiB
references/agents-artifacts.md5.7 KiB
references/agents-background.md3.8 KiB
references/agents-branching.md3.5 KiB
references/agents-custom.md6.7 KiB
references/agents-deployment.md4.1 KiB
references/agents-human-in-the-loop.md8.0 KiB
references/agents-multi-agent.md5.6 KiB
references/agents-sessions.md6.7 KiB
references/agents-state.md7.1 KiB
references/agents.md14.6 KiB
references/dotprompt.md7.9 KiB
references/genkit.md16.2 KiB
references/genkit_anthropic.md1.7 KiB
references/genkit_chrome.md543 B
references/genkit_firebase_ai.md588 B
references/genkit_google_genai.md3.8 KiB
references/genkit_mcp.md4.3 KiB
references/genkit_middleware.md4.7 KiB
references/genkit_openai.md2.8 KiB
references/genkit_shelf.md4.4 KiB
references/genkit_vertexai.md2.6 KiB
references/schemantic.md4.8 KiB

references/agents-human-in-the-loop.md

Agent Human-in-the-Loop / Interrupts

Read agents.md first.

An interrupt pauses an agent mid-turn and hands control back to your code (or a human) — e.g. to approve a sensitive action, collect missing input, or confirm a plan. Internally it's a tool call used as control flow: the interrupt tool never runs to completion on the server; it pauses the turn. You then resume from the exact point it paused.

Two ways to define an interrupt:
ai.defineInterrupt(...): an interrupt-only tool that always pauses and never runs logic. outputSchema describes what the caller supplies on resume.
A normal ai.defineTool whose body returns .interrupt(data): pauses conditionally (e.g. only for large transfers) or after some work.
There is no ctx.interrupt(...); return .interrupt(...) from the tool.

final askUser = ai.defineInterrupt(
  name: 'askUser',
  description: 'Ask the user a clarifying question.',
  inputSchema: .string(), // the question
  outputSchema: .string(), // the answer, supplied on resume
);

Interrupts are orthogonal to persistence — they work the same whether the agent uses a session store or client-managed state. Just resume on the same chat (or, for raw calls, carry the returned state/snapshot back into the resume).

Flow: chat.send(text: ...) → response has res.interrupts → collect human input → chat.resume(respond: [...]).

Define an interrupt (a tool that interrupts)

The example below uses tools that return .interrupt(...), which pauses the turn; its argument is the data shown to the human. Add interrupts (either kind) to the agent's tools.

import 'package:genkit/genkit.dart';
import 'package:genkit/experimental.dart'; // defineAgent, InMemorySessionStore
import 'package:schemantic/schemantic.dart';

import 'genkit.dart';

part 'banking_agent.g.dart';

@Schema()
abstract class $UserApprovalInput {
  @Field(description: 'The action to be approved')
  String get action;
  @Field(description: 'Details about the action')
  String get details;
}

@Schema()
abstract class $TransferMoneyInput {
  double get amount;
  String get toAccount;
}

@Schema()
abstract class $TransferMoneyOutput {
  bool get success;
  String get transactionId;
}

/// Interrupt: always pauses. The caller provides `{ approved, feedback }` on
/// resume. No `outputSchema` — the output comes from the resume call.
final userApproval = ai.defineTool(
  name: 'userApproval',
  description: 'Ask the user for approval before a sensitive action.',
  inputSchema: UserApprovalInput.$schema,
  fn: (input, ctx) async => .interrupt(),
);

/// Executes the transfer. Only reached after the user approves.
final transferMoney = ai.defineTool(
  name: 'transferMoney',
  description: 'Transfer money to a specified account.',
  inputSchema: TransferMoneyInput.$schema,
  outputSchema: TransferMoneyOutput.$schema,
  fn: (input, _) async => .response(TransferMoneyOutput(
    success: true,
    transactionId: 'txn-${DateTime.now().millisecondsSinceEpoch}',
  )),
);

final bankingAgent = ai.defineAgent(
  name: 'bankingAgent',
  system: 'You are a banking assistant. ALWAYS use the userApproval interrupt '
      'to confirm before executing transferMoney.',
  tools: [userApproval, transferMoney],
  use: [retry()],
  store: InMemorySessionStore(),
);

Detect and resume (server-side)

res.interrupts is non-empty when the agent paused. Each AgentInterrupt exposes:

.name — the interrupt's name.
.input — the data the model passed in.
.respond(output) — builder returning a resume entry that supplies the tool's output (without executing it). Does not send.
.restart([payload]) — builder re-issuing the original tool request (retry / let the tool actually run). Pass an optional Map payload; it is nested under metadata.resumed and read back tool-side via ctx.resumed. Does not send.

Resume the same chat with chat.resume(...) / chat.resumeStream(...), passing respond / restart entries directly:

final chat = bankingAgent.chat();
var res = await chat.send(text: 'Transfer \$500 to my savings account.');

final approval =
    res.interrupts.where((i) => i.name == 'userApproval').firstOrNull;
if (approval != null) {
  print(approval.input); // { action, details } — show this to the human

  // Collect the human decision, then resume with the interrupt's output:
  res = await chat.resume(
    respond: [
      approval.respond({'approved': true, 'feedback': 'Looks good'}),
    ],
  );
}
print(res.text); // final confirmation

Streaming variant:

final turn = chat.resumeStream(
  respond: [approval.respond({'approved': true})],
);
await for (final chunk in turn.stream) {
  stdout.write(chunk.text);
}
final res = await turn.response;

You can resume multiple interrupts at once by passing several builders, and mix respond (supply output) with restart (re-run the tool):

await chat.resume(
  respond: [a.respond({'approved': true})],
  restart: [b.restart()],
);

Client-side (browser) interrupts

The same pattern works over HTTP with remoteAgent from package:genkit/client.dart. The client tracks the snapshot, so resuming the same chat continues exactly where it paused.

import 'package:genkit/client.dart';
import 'package:genkit/experimental_client.dart'; // remoteAgent, AgentInterrupt

final agent = remoteAgent(url: '/api/bankingAgent');
final chat = agent.chat();

// 1. Send and detect the pause.
final res = await chat.send(text: 'Transfer \$500 to savings.');
final pending =
    res.interrupts.where((i) => i.name == 'userApproval').firstOrNull;

if (pending != null) {
  // pending.input → { action, details }; render an approval dialog.

  // 2. After the human approves/denies, resume the SAME chat.
  final turn = chat.resumeStream(
    respond: [pending.respond({'approved': true, 'feedback': 'ok'})],
  );
  await for (final chunk in turn.stream) {
    /* render chunk.text */
  }
  final finalRes = await turn.response;
  // If finalRes.interrupts is non-empty, the agent paused again — repeat.
}

Interrupts with tool-approval middleware

The toolApproval middleware turns selected tools into approval interrupts without any custom interrupt code. The middleware gates each tool call itself: a tool is allowed to run only if it is in the approved list or its resumed payload carries { 'tool-approved': true }. To let an approved tool through on resume, restart the paused interrupt with that payload — the .restart(...) builder nests it under metadata.resumed, exactly what the middleware reads:

// `interrupt` is the paused AgentInterrupt from `res.interrupts`.
// Pass this restart entry back when resuming the chat:
await chat.resume(
  restart: [interrupt.restart({'tool-approved': true})],
);

If instead you write your own .interrupt(...) gate inside a tool (as in the coding_agent sample), you inspect the resume payload yourself via the ctx.resumed getter:

final resumed = ctx.resumed; // the payload passed to `.restart(...)`
final isApproved = resumed is Map && resumed['tool-approved'] == true;

Notes & gotchas

No store required. Interrupts work with either a session store or client-managed state.
respond/restart are builders. They return resume entries; they do not send. You still call chat.resume(...).
Resume validation. The server validates each respond/restart entry against the conversation history — always build entries from the interrupt objects in the response, not hand-rolled parts.
Re-pausing. After resuming, the new response may interrupt again; loop until res.interrupts is empty.
UX tip: don't render a model message bubble for an interrupted turn; show the approval UI from interrupt.input instead, then render the model's reply after resuming.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs Node.js

Security

What our scanner noticed· Low risk

Instructions that use the network

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Published by

Google

Publisher

Google (Official)

Source

google/skills · skills/cloud/developing-genkit-dart · commit 2194d3b

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by Google or Google.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download developing-genkit-dart.zip