Genkit Dart 1.0 Is Stable: Build Production Flutter AI Agents Now

Google released Genkit Dart 1.0 on October 8, the first stable, semantically versioned release of its AI framework for Dart, locking the API contract for the core genkit package, the schemantic type-generation library, and the main model plugins. The package had 34,500+ downloads on pub.dev and 160 pub points at launch, and Google's announcement credits feedback from the Dart and Flutter community for shaping the APIs toward production workloads. The release recommends defining AI logic as typed Flows on a Dart server that Flutter apps call via defineRemoteAction, keeping API keys server-side, while stateful agents and A2UI remain under package:genkit/experimental.dart outside the 1.0 stability guarantee.

Flutter developers have been building AI features the hard way — REST calls to a Python backend, manual JSON serialization, no type safety, and model API keys too often exposed in client-side config. Genkit Dart 1.0, released October 8, changes that. Google’s AI framework for Dart has hit its first stable, semantically versioned release. The API contract is locked in. You can ship production apps on top of it today. What the 1.0 Tag Actually Means Before 1.0, Genkit Dart was experimental — Google made no promises about breaking changes. Teams building on it were taking a bet. The 1.0 tag is the stability signal those teams were waiting for: the core genkit package, the schemantic type-generation library, and the main model plugins are now covered by semantic versioning. A breaking change requires a major version bump. That is the commitment. The package already has 34,500+ downloads on pub.dev and 160 pub points at launch. Google’s announcement https://dart.dev/blog/announcing-genkit-dart-1-0 credits “feedback from the Dart and Flutter community” for shaping the APIs toward production workloads. The Architecture: Server Flows, Flutter Client The officially recommended pattern is clean. You define your AI logic as Flows on a Dart server. A Flow is a typed, observable function that gets deployed as an HTTP POST endpoint automatically. Your Flutter app calls those flows using defineRemoteAction , with matching types generated by the schemantic package. API keys stay on the server. Types are shared without manual serialization. Here is what a minimal server flow looks like: final summarizeFlow = ai.defineFlow name: 'summarize', inputSchema: SummarizeInput.$schema, outputSchema: SummarizeOutput.$schema, fn: input, async { final result = await ai.generate model: gemini15Flash, prompt: 'Summarize: ${input.text}', ; return SummarizeOutput summary: result.text ; }, ; await GenkitRouter summarizeFlow .serve port: 3400 ; And the Flutter client call: final summarize = defineRemoteAction<SummarizeInput, SummarizeOutput uri: Uri.parse 'https://your-server/summarize' , inputFromJson: SummarizeInput.fromJson, outputFromJson: SummarizeOutput.fromJson, ; final output = await summarize SummarizeInput text: article ; That is the full stack. One language, shared types, no Python middleman. The Flows documentation https://genkit.dev/docs/dart/flows/ covers streaming, error handling, and the Developer UI — which lets you test flows locally, inspect traces step-by-step, and experiment with prompts before shipping. Three Features Worth Your Attention Unified model API. One generate call works across Gemini, Anthropic Claude, and OpenAI models. Swap the plugin genkit google ai , genkit anthropic , genkit openai , keep the rest of your code. Vendor lock-in is a choice, not a default. Human-in-the-loop interrupts. This is the feature that separates demo agents from agents you can actually ship. Inside a defineTool call, return .interrupt data instead of .response output to pause execution and hand control back to the caller. The Flutter app surfaces a confirmation dialog. The user approves. Generation resumes. Critically, this logic lives inside the tool itself — the model cannot route around it. Production agents that touch real data need this. OpenTelemetry observability. The genkit otel package exports traces, token counts, and latency metrics using the OpenTelemetry GenAI semantic conventions. Your existing observability stack can consume it without proprietary adapters. What Is Still Experimental Two capabilities live under package:genkit/experimental.dart and are explicitly outside the 1.0 stability guarantee: stateful agents conversation persistence via session stores and A2UI generative UI — agents streaming interactive components like date pickers and forms directly into your Flutter app . Both are worth prototyping, but not ready for production bets. Getting Started Add the package and a model plugin: dart pub add genkit dart pub add genkit google ai or genkit anthropic / genkit openai The pub.dev package page https://pub.dev/packages/genkit links to quickstart guides and sample apps. The GitHub repo https://github.com/genkit-ai/genkit-dart has runnable examples covering flows, tools, middleware, and the interrupt pattern. The middleware system — automatic retries, tool approval rules, dynamic skill loading — is covered in a separate announcement on Google Developers Blog https://developers.googleblog.com/announcing-genkit-middleware-intercept-extend-and-harden-your-agentic-apps/ . The Bottom Line Mobile AI has been the gap in the agentic app story. Every major AI framework is Python-first or TypeScript-first. Flutter developers either glued together polyglot stacks or shipped thin wrappers with no real agent logic. Genkit Dart 1.0 is the first stable answer to that problem from Google. The architecture is sound, the type safety is real, and the interrupt mechanism means you can build agents with actual guardrails. The experimental pieces still need time, but the 1.0 core is solid enough to start building on today.