@interview-sdk/adapter-claude

Anthropic Claude provider adapter for @interview-sdk/core, built on the official @anthropic-ai/sdk (Messages API).

Install

npm install @interview-sdk/core @interview-sdk/adapter-claude

Usage

import { ClaudeAdapter } from '@interview-sdk/adapter-claude';

const adapter = new ClaudeAdapter({ apiKey: process.env.ANTHROPIC_API_KEY });
// <InterviewWidget adapter={adapter} mode="client" ... />
// or: new ServerAnswerProcessor({ questions, rubric, adapter })

(AdapterRegistry from @interview-sdk/core is a separate, optional utility for looking an adapter up by string id at runtime — most apps just pass the adapter directly like above.)

ClaudeAdapter accepts an optional model (defaults to claude-opus-4-8) and an optional pre-configured client — useful for testing or for pointing at Claude Platform on AWS / Bedrock / Vertex / Foundry client classes instead of the first-party client.

Behavior

• Splits AIMessage[] into Anthropic's system string param and the messages array (user/assistant only) — matches the security model in @interview-sdk/core, where candidate free text always arrives as its own user-role message.
• Treats a stop_reason: "refusal" (Claude's safety-classifier decline) as an error rather than silently returning empty content.
• Normalizes every Anthropic SDK exception onto the shared Provider*Error taxonomy exported by @interview-sdk/core (ProviderAuthError, ProviderRateLimitError, ProviderOverloadedError, ProviderConnectionError, ProviderTimeoutError, ProviderContextLengthExceededError, ProviderInvalidRequestError), so withRetry and FailoverAdapter from core work without knowing anything Claude-specific.
• Context-length overflow is detected heuristically from the error message text, since Anthropic reports it as a plain 400 invalid_request_error with no distinct error type.
• A deprecated/unknown model id surfaces as Anthropic's 404 NotFoundError and is normalized to ProviderInvalidRequestError — pair this adapter with FailoverAdapter if you want automatic fallback to another provider.

Verified against

@anthropic-ai/[email protected], current as of 2026-07-04 — the Messages API, model catalog, and error taxonomy shift quickly; re-check before assuming this stays accurate for long.