@rudra-js/anthropic

An Anthropic adapter for @rudra-js/core's ComponentProvider.

Anthropic and Claude are trademarks of Anthropic, PBC. This package is an independent adapter and is not affiliated with or endorsed by Anthropic.

Install

npm install @rudra-js/anthropic @rudra-js/core zod@^4

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

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

model defaults to claude-sonnet-5, and thinking defaults to { type: 'disabled' }.

Both of those come from the same constraint. Generation runs on the request path, inside modelTimeoutMs, which core defaults to 1500ms. A model that reasons before it answers won't finish in time. Sonnet 5 reasons by default when thinking is left out, so it's the leaving out that would break the budget, not the choice of model.

Want a different model? Pass model to pin one, and raise modelTimeoutMs to match while you're there. A quick heads-up on thinking: some models reject an explicit { type: 'disabled' } and reason whatever you do. Pass thinking: null for those and we'll send no thinking field at all. Our example shop pins claude-opus-5 and gives modelTimeoutMs a full 60 seconds, which is a demo talking, not a production setting.

maxTokens and baseUrl are optional too.

An identity-linked key needs a workspace

A key you made against your identity rather than inside a workspace belongs to you across several of them. The API can't tell which one a request is acting in, so it answers:

400 invalid_request_error — anthropic-workspace-id is required when
authenticating with an identity-linked API key

Pass the workspace, or create the key from inside a workspace instead:

createAnthropicProvider({
  apiKey: process.env.ANTHROPIC_API_KEY!,
  workspaceId: process.env.ANTHROPIC_WORKSPACE_ID!,
});

fetch is injectable. That's how our test suite stands in for a network call.

The tool schema we send the model is derived from the schema on the ProviderRequest, the same schema @rudra-js/core defines. We don't keep a copy here. A second copy would be a second vocabulary, and then the reconciler enforces one thing while the model is told another.

Data handling

Each generation is one POST to https://api.anthropic.com/v1/messages. Set baseUrl and it goes to that host instead, whether that's a proxy, a gateway, or a region-specific endpoint you have.

What's in that request is listed under What the model sees in the @rudra-js/core README. Read it before you send real shopper traffic. In cohort mode, the default, the request carries no individual. In per-shopper mode it carries that shopper's likes, dislikes, purchases, basket, views and recent searches.

If your shop is in the EU or the UK, you're the one sending personal data to Anthropic, so you'll need a data processing agreement with them plus a transfer mechanism for the data leaving your region. Your shop is Anthropic's customer. This package is a piece of code in the middle, and it isn't a party to anything.

When the API answers with an error, the Error we throw carries the status code and the vendor's error category, like anthropic responded 400 (invalid_request_error). We leave the vendor's own message out. It quotes the request back, and for this framework the request can hold a shopper's search terms, which an adopter's console.error(error) would then write to a log. A test in anthropic-provider.test.ts puts a search term in that message and asserts it does not reach the thrown error.

Licence

MIT