createAnthropicProvider

import { createAnthropicProvider } from '@rudra-js/anthropic';

Call createAnthropicProvider to connect the generator to Claude through the Anthropic API.

createAnthropicProvider(options: AnthropicProviderOptions) => ComponentProvider

Reference

`createAnthropicProvider(options)`

Parameters

options: Your API key, and settings for the calls.
- apiKey: Your Anthropic API key. No rudra-js package reads the environment, so pass it from wherever you keep secrets.
- modeloptional: The Claude model to call. Defaults to 'claude-sonnet-5'.
- maxTokensoptional: The most tokens the model may write in one answer. Defaults to 8192.
- baseUrloptional: Where the calls go. Defaults to https://api.anthropic.com. Set it to send them through your own proxy, a gateway or a regional endpoint. A gateway someone else runs sees every request and your apiKey.
- workspaceIdoptional: The Anthropic workspace the calls act in, sent as the anthropic-workspace-id header. An identity-linked API key needs it.
- thinkingoptional: Whether the model reasons before it answers. Defaults to { type: 'disabled' }. Pass { type: 'adaptive' } to let the model decide. Pass null to send no thinking field, for a model that rejects the default.
- fetchoptional: The fetch the calls go through. Defaults to the global fetch. Pass your own to test without the network.

Returns

A ComponentProvider to pass to createComponentGenerator as provider. Its name is 'anthropic'. Its model is the model it calls.

Caveats

When you pick another model or turn thinking on, raise modelTimeoutMs to match. A model that reasons before it answers won't finish inside the default. Each request that runs out of time gets the block core builds without a model.
Read Running it in a regulated shop before you send real shopper traffic. It lists what each call carries.
When the API answers with an error, the thrown error holds only the status and the error type, such as anthropic responded 400 (invalid_request_error). It leaves out the API's own message, which can quote a shopper's searches.

Usage

Picking another Claude model

Pass model, and give the generator a modelTimeoutMs that fits it.

import { createAnthropicProvider } from '@rudra-js/anthropic';
import { createComponentGenerator } from '@rudra-js/core';

export const generator = createComponentGenerator({
  provider: createAnthropicProvider({
    apiKey: process.env.ANTHROPIC_API_KEY!,
    model: 'claude-opus-5',
  }),
  modelTimeoutMs: 5_000,
});

With Claude Opus 5.5, Sonnet 5.5, Fable 5.1 or Mythos 5.1, also pass thinking: null.

Time a few calls to the model you picked, then set modelTimeoutMs above them.

Log event.source to check that the model answered.

Troubleshooting

400 … anthropic-workspace-id is required

The API answers this way when an identity-linked key doesn't say which workspace a call is for. In event.error it shows as anthropic responded 400 (invalid_request_error).

A key you made against your identity, rather than inside a workspace, can act in several workspaces. Pass workspaceId to pick one, or create the key inside a workspace.

import { createAnthropicProvider } from '@rudra-js/anthropic';
import { createComponentGenerator } from '@rudra-js/core';

export const generator = createComponentGenerator({
  provider: createAnthropicProvider({
    apiKey: process.env.ANTHROPIC_API_KEY!,
    workspaceId: process.env.ANTHROPIC_WORKSPACE_ID!,
  }),
});

`source` is always `fallback`

Log event.degradedReason and event.error, then start with the generator's troubleshooting.

A model that rejects thinking: { type: 'disabled' } fails every call. Claude Opus 5.5, Sonnet 5.5, Fable 5.1 and Mythos 5.1 do. Pass thinking: null with them. They reason before they answer, so raise modelTimeoutMs too.

anthropic returned no emit_component_spec tool use means the model answered in plain text instead of calling the tool. That request gets the block core builds without a model.

anthropic stopped at the max_tokens budget means the answer ran past maxTokens. Raise it.

Source on GitHub @rudra-js/anthropic on npm