developing-genkit-dart

Use ai.defineTool with a name, description, inputSchema, and an async fn implementing the logic. Input schemas are defined with the schemantic library using @Schema() abstract classes prefixed with $, then passed as YourClass.$schema.

Call ai.generateStream with a model reference such as googleAI.gemini('gemini-2.5-flash') and a prompt, then iterate over the returned stream with await for to read each chunk's text.

Yes, the genkit_openai plugin accepts a custom baseUrl and CustomModelDefinition entries, so providers like Groq, DeepSeek, and Together AI can be used by overriding the endpoint and declaring model capabilities.

Use the genkit_mcp package's defineMcpHost to connect to one or more MCP servers and automatically register their tools. Tools can then be referenced in ai.generate via wildcard names like 'my-host:tool/fs/*' or exact tool names.

Schemantic requires running dart run build_runner build to generate the .g.dart files for @Schema() classes. Missing part directives or ungenerated files cause compile errors, so verify the part statement matches the filename and rerun the builder.

Use the genkit_shelf plugin's startFlowServer for a standalone server, or mount flows onto an existing Shelf Router with shelfHandler. Deployed flows can then be called from Dart or JavaScript clients using defineRemoteAction or the genkit npm client.