Skip to main content
Genkit is designed around a flexible, plugin-based architecture that makes it easy to build AI applications with any model provider, telemetry backend, or vector store.

​
Architecture Overview

At the heart of Genkit is a Registry that manages all actions (flows, tools, models, prompts, etc.) and coordinates between your application code, plugins, and observability systems. 
┌─────────────────────────────────────────────────────────────────────────────┐
│                         Your Application                                    │
│                                                                             │
│   from genkit import Genkit                                                 │
│   ai = Genkit(plugins=[GoogleGenAI()], model=gemini_2_0_flash)             │
│                                                                             │
│   ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌──────────┐  ┌───────────┐     │
│   │  Flows  │  │  Tools  │  │ Prompts │  │ Embedders│  │ Retrievers│     │
│   │@ai.flow │  │@ai.tool │  │.prompt  │  │ai.embed()│  │ai.retrieve│     │
│   └────┬────┘  └────┬────┘  └────┬────┘  └────┬─────┘  └─────┬─────┘     │
│        │            │            │             │             │              │
│        └────────────┴────────────┴─────────────┴─────────────┘              │
│                                  │                                          │
│                          ┌───────▼────────┐                                 │
│                          │    Registry     │                                 │
│                          │  (all actions)  │                                 │
│                          └───────┬────────┘                                 │
│                                  │                                          │
│             ┌────────────────────┼────────────────────┐                     │
│             ▼                    ▼                    ▼                     │
│      ┌────────────┐     ┌──────────────┐     ┌──────────────┐             │
│      │   Plugins   │     │   Tracing     │     │  Reflection  │             │
│      │ (providers) │     │ (OpenTelemetry│     │  API (DevUI) │             │
│      └────────────┘     └──────────────┘     └──────────────┘             │
└─────────────────────────────────────────────────────────────────────────────┘

​
Core Components

​
Registry

The Registry is Genkit’s central nervous system. It:
  • Stores all actions: Flows, tools, models, prompts, embedders, and retrievers
  • Resolves actions by name: Looks up actions when you reference them (e.g., model='googleai/gemini-2.0-flash')
  • Manages plugins: Coordinates plugin initialization and action resolution
  • Powers the Dev UI: Provides the Reflection API that lets the Developer UI discover and run your actions

​
Actions

Everything in Genkit is an Action - a unit of work with:
  • Name and kind: Identifies the action type (flow, tool, model, etc.)
  • Input/output schemas: Type-safe contracts using Zod (JS) or Pydantic (Python)
  • Execution function: The actual code that runs
  • Automatic tracing: Every action execution is recorded for observability
// JavaScript
export const myFlow = ai.defineFlow(
  { name: 'myFlow', inputSchema: z.string(), outputSchema: z.string() },
  async (input) => {
    return `Hello, ${input}!`;
  }
);

# Python
@ai.flow()
async def my_flow(input: str) -> str:
    return f"Hello, {input}!"

// Go
func MyFlow(ctx context.Context, input string) (string, error) {
    return fmt.Sprintf("Hello, %s!", input), nil
}

flow := core.DefineFlow("myFlow", MyFlow)

​
Plugins

Plugins extend Genkit’s capabilities. Each plugin:
  • Implements a common interface: Plugin base class with init(), resolve(), and list_actions() methods
  • Loads lazily: Only initialized when first used, not at startup
  • Registers actions: Can pre-register actions during init() or resolve them on-demand
  • Has a namespace: Prevents naming conflicts (e.g., googleai/gemini-2.0-flash)
┌─────────────────────────────────────────────────────────────────────┐
│                  Plugin (Abstract Base Class)                       │
│                  genkit.core.plugin.Plugin                          │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  name: str                                                          │
│  ─────────────────────────────────────                              │
│  Plugin namespace (e.g., 'googleai', 'anthropic', 'ollama')        │
│                                                                     │
│  async init() → list[Action]            ← called once per plugin   │
│  ─────────────────────────────────────                              │
│  One-time initialization; returns actions to pre-register.          │
│  Called lazily on first action resolution, NOT at registration.     │
│                                                                     │
│  async resolve(kind, name) → Action?    ← called per lookup        │
│  ─────────────────────────────────────                              │
│  Resolve a single action by kind and name. Returns None if         │
│  this plugin doesn't handle the requested action.                   │
│                                                                     │
│  async list_actions() → list[ActionMetadata]  ← for Dev UI         │
│  ─────────────────────────────────────                              │
│  Advertise available actions without heavy initialization.          │
│  Called by the Reflection API for DevUI action discovery.           │
└─────────────────────────────────────────────────────────────────────┘

​
Plugin Lifecycle

Plugins go through four phases: 
┌─────────────────────────────────────────────────────────────────────┐
│                      Plugin Lifecycle                               │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  Phase 1: REGISTRATION (at Genkit startup)                         │
│  ─────────                                                         │
│  ai = Genkit(plugins=[GoogleAI(), Anthropic()])                    │
│       │                                                            │
│       ├─► registry.register_plugin(GoogleAI())                     │
│       └─► registry.register_plugin(Anthropic())                    │
│           │                                                        │
│           ▼                                                        │
│       ┌────────────────────────────────────────┐                   │
│       │  Registry._plugins                      │                   │
│       │  ┌────────────────┬───────────────────┐ │                   │
│       │  │ "googleai"     │ GoogleAI instance │ │                   │
│       │  │ "anthropic"    │ Anthropic instance│ │                   │
│       │  └────────────────┴───────────────────┘ │                   │
│       └────────────────────────────────────────┘                   │
│                                                                     │
│  Phase 2: LAZY INIT (on first action resolution)                   │
│  ──────────                                                        │
│  await ai.generate(model='googleai/gemini-2.0-flash', ...)        │
│       │                                                            │
│       ▼                                                            │
│  registry._ensure_plugin_initialized('googleai')                   │
│       │                                                            │
│       ▼                                                            │
│  actions = await plugin.init()    ← called exactly once            │
│       │                            (subsequent calls are no-ops)   │
│       ▼                                                            │
│  for action in actions:                                            │
│      registry.register_action_instance(action)                     │
│                                                                     │
│  Phase 3: ACTION RESOLUTION (on each usage)                        │
│  ─────────────────────                                             │
│  await plugin.resolve(ActionKind.MODEL, 'googleai/gemini-2.0-flash')│
│       │                                                            │
│       ▼                                                            │
│  Action instance returned and cached in registry                   │
│                                                                     │
│  Phase 4: ACTION DISCOVERY (for Dev UI)                            │
│  ──────────────────                                                │
│  await plugin.list_actions()                                       │
│       │                                                            │
│       ▼                                                            │
│  ActionMetadata[] returned to Reflection API                       │
│  (does NOT trigger init — must be fast and safe)                   │
└─────────────────────────────────────────────────────────────────────┘

​
Action Resolution

When you reference an action by name, the Registry uses a multi-step resolution algorithm: 
ai.generate(model="googleai/gemini-2.0-flash")
     │
     ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 1: CACHE HIT?                                              │
│  Is "googleai/gemini-2.0-flash" already in registry._entries?    │
│     ├── YES → return cached Action (fast path)                   │
│     └── NO  → continue to Step 2                                 │
├──────────────────────────────────────────────────────────────────┤
│  Step 2: NAMESPACED or UNPREFIXED?                               │
│  Does the name contain "/"?                                      │
│     │                                                            │
│     ├── YES ("googleai/gemini-2.0-flash")                        │
│     │    ├── Find plugin "googleai"                               │
│     │    ├── await _ensure_plugin_initialized("googleai")         │
│     │    ├── Check cache again (init may have registered it)     │
│     │    └── await plugin.resolve(MODEL, "googleai/gemini-2.0")  │
│     │                                                            │
│     └── NO ("gemini-2.0-flash")                                  │
│          ├── Try ALL plugins                                      │
│          ├── If 1 match  → use it                                 │
│          ├── If 2+ match → ValueError (ambiguous)                │
│          └── If 0 match  → continue to Step 3                    │
├──────────────────────────────────────────────────────────────────┤
│  Step 3: DYNAMIC ACTION PROVIDERS (fallback)                     │
│  Try registered Dynamic Action Providers (e.g., MCP servers)     │
│     ├── Found → register and return                              │
│     └── Not found → return None                                  │
└──────────────────────────────────────────────────────────────────┘

​
Tracing and Observability

Genkit automatically instruments all actions with OpenTelemetry tracing: 
ai.generate(prompt="Tell me a joke")
     │
     ▼
┌──────────┐    ┌───────────┐    ┌──────────┐    ┌──────────┐
│  1. Flow  │───►│ 2. Model  │───►│ 3. Tool? │───►│ 4. Model │
│  starts   │    │  called   │    │  (if the │    │ responds │
│  tracing  │    │  (Gemini, │    │  model   │    │  with    │
│           │    │  Claude,  │    │  decides │    │  final   │
│           │    │  etc.)    │    │  to use  │    │  answer  │
│           │    │           │    │  one)    │    │          │
└──────────┘    └───────────┘    └──────────┘    └──────────┘
     │                │               │               │
     └────────────────┴───────────────┴───────────────┘
                              │
                  ┌───────────▼───────────┐
                  │   Trace (every step   │
                  │   recorded for DevUI, │
                  │   Cloud Trace, etc.)  │
                  └──────────────────────┘
Every action execution creates a span with:
  • Input and output data
  • Execution duration
  • Error information (if any)
  • Custom metadata

​
Multi-Language Support

Genkit maintains feature parity across three languages:
LanguagePackageStatus
JavaScript/TypeScriptgenkitPrimary SDK
PythongenkitFeature parity with JS
Gogithub.com/firebase/genkit/go/coreProduction-ready
The architecture is consistent across all three:
  • Same plugin model
  • Same action types
  • Same tracing integration
  • Same Dev UI support

​
Next Steps

  • Learn about Flows - the building blocks of Genkit applications
  • Explore Plugins - how to extend Genkit’s capabilities
  • Understand Observability - tracing and monitoring your AI apps

Build docs developers (and LLMs) love

Get started for freeTalk to us