Creating Plugins

Genkit plugins are modular extensions that add new capabilities to the framework. You can create custom plugins to integrate new model providers, vector stores, or any other functionality.

Plugin Architecture

A Genkit plugin implements the PluginProvider interface:

export interface PluginProvider {
  name: string;
  initializer: () =>
    | InitializedPlugin
    | void
    | Promise<InitializedPlugin | void>;
  resolver?: (action: ActionType, target: string) => Promise<void>;
  listActions?: () => Promise<ActionMetadata[]>;
}

export interface InitializedPlugin {
  models?: Action<z.ZodTypeAny, z.ZodTypeAny>[];
  retrievers?: Action<z.ZodTypeAny, z.ZodTypeAny>[];
  embedders?: Action<z.ZodTypeAny, z.ZodTypeAny>[];
  indexers?: Action<z.ZodTypeAny, z.ZodTypeAny>[];
  evaluators?: Action<z.ZodTypeAny, z.ZodTypeAny>[];
}

Plugin Types

Plugin API v1 (Legacy)

The original plugin API uses genkitPlugin:

import { genkitPlugin, type GenkitPlugin } from 'genkit/plugin';
import type { Genkit } from 'genkit';

export function myPlugin(options?: MyPluginOptions): GenkitPlugin {
  return genkitPlugin(
    'my-plugin',
    async (ai: Genkit) => {
      // Register actions with the Genkit instance
      ai.defineModel(/* ... */);
      ai.defineRetriever(/* ... */);
    },
    async (ai, actionType, actionName) => {
      // Optional: Resolve actions dynamically
    },
    async () => {
      // Optional: List available actions
      return [];
    }
  );
}

Plugin API v2 (Recommended)

The new plugin API supports direct model usage without Genkit instance:

import { genkitPluginV2, type GenkitPluginV2 } from 'genkit/plugin';
import { ModelAction } from 'genkit/model';
import { ActionType } from 'genkit/registry';

function myPlugin(options?: MyPluginOptions): GenkitPluginV2 {
  return genkitPluginV2({
    name: 'my-plugin',
    init: async () => {
      // Return actions to register
      const actions: ModelAction[] = [];
      for (const modelName of AVAILABLE_MODELS) {
        actions.push(createModel(modelName));
      }
      return actions;
    },
    resolve: (actionType: ActionType, name: string) => {
      // Dynamically create actions on demand
      if (actionType === 'model') {
        return createModel(name);
      }
      return undefined;
    },
    list: async () => {
      // List available actions for discovery
      return await listModels();
    },
  });
}

Creating a Model Provider Plugin

Basic Model Plugin

import {
  modelRef,
  type ModelReference,
  type GenerateRequest,
  type GenerateResponseData,
} from 'genkit/model';
import { genkitPluginV2, type GenkitPluginV2 } from 'genkit/plugin';
import { z } from 'genkit';

interface MyProviderOptions {
  apiKey?: string;
}

const MyProviderConfigSchema = z.object({
  temperature: z.number().min(0).max(2).optional(),
  maxTokens: z.number().optional(),
});

function myProviderPlugin(options?: MyProviderOptions): GenkitPluginV2 {
  const apiKey = options?.apiKey || process.env.MY_PROVIDER_API_KEY;
  
  if (!apiKey) {
    throw new Error('API key required');
  }

  return genkitPluginV2({
    name: 'my-provider',
    init: async () => {
      return [
        createModel('my-model-1', apiKey),
        createModel('my-model-2', apiKey),
      ];
    },
  });
}

function createModel(name: string, apiKey: string) {
  return {
    name: `my-provider/${name}`,
    label: `My Provider - ${name}`,
    configSchema: MyProviderConfigSchema,
    info: {
      supports: {
        multiturn: true,
        media: false,
        tools: true,
        systemRole: true,
      },
    },
    async generate(
      input: GenerateRequest,
      streamingCallback?: (chunk: any) => void
    ): Promise<GenerateResponseData> {
      // Call your model API
      const response = await fetch('https://api.myprovider.com/generate', {
        method: 'POST',
        headers: {
          'Authorization': `Bearer ${apiKey}`,
          'Content-Type': 'application/json',
        },
        body: JSON.stringify({
          messages: input.messages,
          config: input.config,
        }),
      });

      const data = await response.json();

      return {
        message: {
          role: 'model',
          content: [{ text: data.output }],
        },
        usage: {
          inputTokens: data.usage.input_tokens,
          outputTokens: data.usage.output_tokens,
        },
        finishReason: 'stop',
      };
    },
  };
}

// Export with model reference helper
export const myProvider = myProviderPlugin as typeof myProviderPlugin & {
  model: (name: string, config?: any) => ModelReference<typeof MyProviderConfigSchema>;
};

myProvider.model = (name: string, config?: any) => {
  return modelRef({
    name: `my-provider/${name}`,
    config,
    configSchema: MyProviderConfigSchema,
  });
};

Creating a Vector Store Plugin

Retriever and Indexer Plugin

import {
  retrieverRef,
  indexerRef,
  type Document,
  type Genkit,
} from 'genkit';
import { genkitPlugin, type GenkitPlugin } from 'genkit/plugin';
import { z } from 'genkit';
import type { EmbedderArgument } from 'genkit/embedder';

interface VectorStoreConfig {
  collectionName: string;
  embedder: EmbedderArgument<any>;
  connectionUrl?: string;
}

const RetrieverOptionsSchema = z.object({
  k: z.number().optional(),
  filter: z.record(z.any()).optional(),
});

export function myVectorStore(
  configs: VectorStoreConfig[]
): GenkitPlugin {
  return genkitPlugin('my-vectorstore', async (ai: Genkit) => {
    configs.forEach(config => {
      defineRetriever(ai, config);
      defineIndexer(ai, config);
    });
  });
}

function defineRetriever(
  ai: Genkit,
  config: VectorStoreConfig
) {
  return ai.defineRetriever(
    {
      name: `my-vectorstore/${config.collectionName}`,
      configSchema: RetrieverOptionsSchema,
    },
    async (query, options) => {
      // 1. Embed the query
      const embeddings = await ai.embed({
        embedder: config.embedder,
        content: query,
      });

      // 2. Search your vector store
      const results = await searchVectorStore({
        collection: config.collectionName,
        vector: embeddings[0].embedding,
        limit: options?.k || 10,
        filter: options?.filter,
      });

      // 3. Return documents
      return {
        documents: results.map(r => ({
          content: [{ text: r.content }],
          metadata: r.metadata,
        })),
      };
    }
  );
}

function defineIndexer(
  ai: Genkit,
  config: VectorStoreConfig
) {
  return ai.defineIndexer(
    {
      name: `my-vectorstore/${config.collectionName}`,
    },
    async (docs) => {
      // 1. Embed all documents
      const embeddings = await Promise.all(
        docs.map(doc =>
          ai.embed({
            embedder: config.embedder,
            content: doc,
          })
        )
      );

      // 2. Prepare records for insertion
      const records = embeddings.flatMap((embeddingList, docIndex) => {
        const doc = docs[docIndex];
        return embeddingList.map(embedding => ({
          id: generateId(),
          vector: embedding.embedding,
          content: doc.text,
          metadata: doc.metadata,
        }));
      });

      // 3. Insert into vector store
      await insertVectors({
        collection: config.collectionName,
        records: records,
      });
    }
  );
}

// Helper reference functions
export const myVectorStoreRetrieverRef = (params: {
  collectionName: string;
}) => {
  return retrieverRef({
    name: `my-vectorstore/${params.collectionName}`,
    configSchema: RetrieverOptionsSchema,
  });
};

export const myVectorStoreIndexerRef = (params: {
  collectionName: string;
}) => {
  return indexerRef({
    name: `my-vectorstore/${params.collectionName}`,
  });
};

// Mock implementations
async function searchVectorStore(params: any) {
  // Implement your vector search
  return [];
}

async function insertVectors(params: any) {
  // Implement your vector insertion
}

function generateId(): string {
  return Math.random().toString(36).substring(7);
}

Creating an Embedder Plugin

import {
  embedderRef,
  type EmbedderReference,
  type Genkit,
} from 'genkit';
import { genkitPlugin, type GenkitPlugin } from 'genkit/plugin';
import { z } from 'genkit';

interface EmbedderConfig {
  name: string;
  dimensions: number;
  apiKey?: string;
}

export function myEmbedderPlugin(
  configs: EmbedderConfig[]
): GenkitPlugin {
  return genkitPlugin('my-embedder', async (ai: Genkit) => {
    configs.forEach(config => {
      ai.defineEmbedder(
        {
          name: `my-embedder/${config.name}`,
          info: {
            dimensions: config.dimensions,
            label: `My Embedder - ${config.name}`,
          },
        },
        async (input) => {
          // Call your embedding API
          const response = await fetch('https://api.myembedder.com/embed', {
            method: 'POST',
            headers: {
              'Authorization': `Bearer ${config.apiKey}`,
              'Content-Type': 'application/json',
            },
            body: JSON.stringify({
              texts: input.content.map(c => c.text),
            }),
          });

          const data = await response.json();

          return {
            embeddings: data.embeddings.map((embedding: number[]) => ({
              embedding,
            })),
          };
        }
      );
    });
  });
}

// Helper reference function
export const myEmbedderRef = (params: {
  name: string;
}) => {
  return embedderRef({
    name: `my-embedder/${params.name}`,
  });
};

Plugin Lifecycle

Initialization Phase

User configures plugin - When calling genkit({ plugins: [...] })
Plugin initializer runs - The init function is called
Actions registered - Models, retrievers, etc. are registered with Genkit

Resolution Phase (Optional)

User references action - e.g., myProvider.model('unknown-model')
Resolver called - The resolve function tries to create the action dynamically
Action returned - If successful, action is used immediately

Listing Phase (Optional)

Developer UI queries - Or programmatic discovery
List function called - Returns metadata about available actions
Actions displayed - User can see what’s available

Best Practices

1. Use TypeScript

// Define clear interfaces
interface MyPluginOptions {
  apiKey?: string;
  baseUrl?: string;
}

// Export typed plugin
export type MyPlugin = {
  (options?: MyPluginOptions): GenkitPluginV2;
  model: (name: string) => ModelReference;
};

2. Handle Authentication

function myPlugin(options?: MyPluginOptions): GenkitPluginV2 {
  const apiKey = options?.apiKey || process.env.MY_API_KEY;
  
  if (!apiKey) {
    throw new Error(
      'API key required. Pass apiKey option or set MY_API_KEY env var.'
    );
  }
  
  // Use apiKey in requests
}

3. Provide Model References

// Export helper for type-safe model references
export const myProvider = myProviderPlugin as typeof myProviderPlugin & {
  model: (name: string) => ModelReference;
};

myProvider.model = (name: string) => {
  return modelRef({
    name: `my-provider/${name}`,
    configSchema: MyConfigSchema,
  });
};

4. Support Streaming

async generate(
  input: GenerateRequest,
  streamingCallback?: (chunk: any) => void
): Promise<GenerateResponseData> {
  if (streamingCallback) {
    // Handle streaming
    const stream = await fetchStream(input);
    for await (const chunk of stream) {
      streamingCallback({
        index: 0,
        content: [{ text: chunk.text }],
      });
    }
  }
  
  // Return final response
}

5. Error Handling

try {
  const response = await fetch(url, options);
  
  if (!response.ok) {
    throw new Error(`API error: ${response.status} ${response.statusText}`);
  }
  
  return await response.json();
} catch (error) {
  throw new Error(`Failed to call API: ${error.message}`);
}

6. Document Configuration

/**
 * My custom AI provider plugin.
 * 
 * @param options - Plugin configuration
 * @param options.apiKey - API key (or set MY_API_KEY env var)
 * @param options.baseUrl - Optional custom API base URL
 * 
 * @example
 * ```typescript
 * import { myProvider } from 'my-genkit-plugin';
 * 
 * const ai = genkit({
 *   plugins: [
 *     myProvider({ apiKey: 'key' }),
 *   ],
 * });
 * ```
 */
export function myProvider(options?: MyProviderOptions): GenkitPluginV2 {
  // ...
}

Testing Plugins

import { describe, it, expect } from 'vitest';
import { genkit } from 'genkit';
import { myProvider } from './my-plugin';

describe('myProvider plugin', () => {
  it('should register models', async () => {
    const ai = genkit({
      plugins: [myProvider({ apiKey: 'test' })],
    });

    // Test model is available
    const response = await ai.generate({
      model: myProvider.model('test-model'),
      prompt: 'Hello',
    });

    expect(response.text()).toBeTruthy();
  });

  it('should handle errors', async () => {
    expect(() => {
      myProvider({ apiKey: undefined });
    }).toThrow('API key required');
  });
});

Publishing Plugins

See Publishing Plugins for details on publishing your plugin to npm.

Examples

Study these official plugins for reference:

Anthropic Plugin - Model provider with Plugin API v2
Ollama Plugin - Model provider with custom config
Chroma Plugin - Vector store with retriever/indexer
Pinecone Plugin - Vector store with namespaces

Official Plugins

Plugin Development

Creating Plugins

Creating Plugins

Plugin Architecture

Plugin Types

Plugin API v1 (Legacy)

Plugin API v2 (Recommended)

Creating a Model Provider Plugin

Basic Model Plugin

Creating a Vector Store Plugin

Retriever and Indexer Plugin

Creating an Embedder Plugin

Plugin Lifecycle

Initialization Phase

Resolution Phase (Optional)

Listing Phase (Optional)

Best Practices

1. Use TypeScript

2. Handle Authentication

3. Provide Model References

4. Support Streaming

5. Error Handling

6. Document Configuration

Testing Plugins

Publishing Plugins

Examples

Links

Build docs developers (and LLMs) love

Official Plugins

Plugin Development

​Creating Plugins

​Plugin Architecture

​Plugin Types

​Plugin API v1 (Legacy)

​Plugin API v2 (Recommended)

​Creating a Model Provider Plugin

​Basic Model Plugin

​Creating a Vector Store Plugin

​Retriever and Indexer Plugin

​Creating an Embedder Plugin

​Plugin Lifecycle

​Initialization Phase

​Resolution Phase (Optional)

​Listing Phase (Optional)

​Best Practices

​1. Use TypeScript

​2. Handle Authentication

​3. Provide Model References

​4. Support Streaming

​5. Error Handling

​6. Document Configuration

​Testing Plugins

​Publishing Plugins

​Examples

​Links

Build docs developers (and LLMs) love

Creating Plugins

Plugin Architecture

Plugin Types

Plugin API v1 (Legacy)

Plugin API v2 (Recommended)

Creating a Model Provider Plugin

Basic Model Plugin

Creating a Vector Store Plugin

Retriever and Indexer Plugin

Creating an Embedder Plugin

Plugin Lifecycle

Initialization Phase

Resolution Phase (Optional)

Listing Phase (Optional)

Best Practices

1. Use TypeScript

2. Handle Authentication

3. Provide Model References

4. Support Streaming

5. Error Handling

6. Document Configuration

Testing Plugins

Publishing Plugins

Examples

Links