effect-ai-provider

Create a client layer with ProviderClient.layerConfig using Config.redacted for the API key, then provide FetchHttpClient.layer. Use ProviderLanguageModel.model(modelId) to get a Model for Effect.provide or ExecutionPlan, or .layer({ model }) for a raw Layer.

Use ExecutionPlan.make with ordered entries specifying provide (a Model) and attempts (retry count). Call plan.captureRequirements inside Layer.effect to satisfy client requirements, then apply it with Effect.withExecutionPlan on your effect.

No @effect/ai-google or @effect/ai-amazon-bedrock packages exist. Use @effect/ai-openrouter with provider-prefixed model IDs like anthropic/claude-sonnet-4 to access Google, Bedrock, and other models through the OpenRouter proxy.

The most common cause is a missing FetchHttpClient.layer, since all provider clients require an HttpClient. Also verify API keys use Config.redacted rather than hardcoded strings, and that client layers are provided to the language model layer.

Create a session with Chat.fromPrompt using Prompt.setSystem for the system message, then call session.generateText which maintains history automatically. Access history via Ref.get(session.history), persist with session.exportJson, and restore with Chat.fromJson.

Yes, use the provider's withConfigOverride dual API, such as AnthropicLanguageModel.withConfigOverride({ temperature: 0.7, max_tokens: 4096 }), piped onto any effect using the LanguageModel. OpenAI overrides also support reasoning, text verbosity, and strictJsonSchema.