@llm-ports/adapter-anthropic

Direct adapter for the Anthropic SDK. Implements LLMPort for Claude models. Does NOT implement EmbeddingsPort — Anthropic ships no embedding models.

Install

pnpm add @llm-ports/core @llm-ports/adapter-anthropic @anthropic-ai/sdk zod

Configure

import { createRegistryFromEnv } from "@llm-ports/core";
import { createAnthropicAdapter } from "@llm-ports/adapter-anthropic";

const registry = createRegistryFromEnv({
  adapters: {
    anthropic: createAnthropicAdapter({
      apiKey: process.env.ANTHROPIC_API_KEY!,
    }),
  },
});

export const llm = registry.getPort();

.env:

LLM_PROVIDER_FAST=anthropic|claude-haiku-4-5|cost:5/day
LLM_PROVIDER_PREMIUM=anthropic|claude-sonnet-4-6-20250514|cost:50/day
LLM_TASK_ROUTE_TRIAGE=fast,premium

Adapter options

interface AnthropicAdapterOptions {
  apiKey: string;
  baseURL?: string;                 // typically only useful for testing
  fetch?: typeof fetch;             // inject a custom fetch (tests, proxies)
  validationStrategy?: ValidationStrategy;
  pricingOverrides?: Record<string, ModelPricing>;
  imageSizeLimitBytes?: number;     // default 5 MB (Anthropic's per-image limit)
  dangerouslyAllowBrowser?: boolean; // alpha.9; opt in to browser execution
  onRetry?: OnRetry;                // observability hook for retries
}

dangerouslyAllowBrowser (alpha.9)

The Anthropic SDK refuses to construct in a browser environment unless dangerouslyAllowBrowser: true is passed explicitly. When enabled, the SDK auto-adds the anthropic-dangerous-direct-browser-access header on every request. Set this option only when the API key is NOT a long-lived secret: short-lived proxy tokens, BYO-key UIs where the end user supplies their own key, or trusted internal tools running behind auth. For server-side proxy patterns where the secret stays on the server, leave it unset.

const adapter = createAnthropicAdapter({
  apiKey: ephemeralUserKey,
  dangerouslyAllowBrowser: true,
});

Claude 4.5+ temperature deprecation (alpha.10)

Anthropic deprecated the temperature parameter on the Claude 4 Opus + Sonnet reasoning family. The static catalog seeds temperatureLocked: true BEFORE the first call against any claude-opus-4-N or claude-sonnet-4-N model (N >= 5), so:

• Non-streaming methods skip the wasted "send temperature, get 400, retry without" round-trip.
• Streaming methods (streamText, streamStructured) work at all — they can't mid-stream retry, so the catalog hit is the ONLY thing that prevents a hard 400.

Bare claude-opus-4 (the original 4.0 release) and the Claude Haiku 4-5 family still accept temperature and are passed through unchanged. The Anthropic adapter's applyCapabilityFilter strips temperature from the request body when the model is catalog-marked or has been learned at runtime.

Bundled pricing

Model	Input/1M	Output/1M	Cache read	Cache write
claude-opus-4-7	$15.00	$75.00	$1.50	$18.75
claude-opus-4	$15.00	$75.00	$1.50	$18.75
claude-sonnet-4-6-20250514	$3.00	$15.00	$0.30	$3.75
claude-sonnet-4-5	$3.00	$15.00	$0.30	$3.75
claude-haiku-4-5	$0.80	$4.00	$0.08	$1.00

Source: anthropic.com/pricing. Verified 2026-05-26.

Override per model via pricingOverrides if prices change between releases:

createAnthropicAdapter({
  apiKey: process.env.ANTHROPIC_API_KEY!,
  pricingOverrides: {
    "claude-sonnet-4-6-20250514": { inputPer1M: 3.5, outputPer1M: 17.5 },
  },
});

Supported features

Feature	Status
generateText	✓
generateStructured (Zod schemas)	✓ (prompted JSON + retry-with-feedback)
streamText	✓
streamStructured	✓ (best-effort partial parse)
runAgent (multi-turn tool use)	✓
Vision input — base64 images	✓
Vision input — URL images	✓
Audio input	✗ (Anthropic chat doesn't support audio)
Prompt caching	✓ native; cost reflects cacheReadPer1M rate
AbortSignal cancellation	✓ entry + in-flight (alpha.6)
listModels()	✓ (alpha.9; via direct fetch to /v1/models)
dangerouslyAllowBrowser opt-in	✓ (alpha.9)
Claude 4.5+ temperature auto-strip	✓ catalog + runtime learning (alpha.3, expanded alpha.10)
Embeddings	✗ (Anthropic ships no embedding models)

Content blocks supported

text, image (base64 + URL), tool_use, tool_result. The adapter throws ContentBlockUnsupportedError for audio blocks.

Cancellation

Full AbortSignal support shipped in 0.1.0-alpha.6. The signal is threaded into both client.messages.create (non-streaming) and client.messages.stream, so controller.abort() cancels the in-flight HTTP request. runAgent also re-checks the signal between steps so cancellation propagates mid-loop. See the Cancellation guide.

Structured output from a JSON Schema

Since 0.1.0-alpha.35, generateStructured accepts jsonSchema in place of schema, for a caller that already holds a JSON Schema rather than a Zod one:

await llm.generateStructured<Triage>({ taskType: "classify", messages, jsonSchema });

Exactly one of the two is required. What the JSON Schema path gives up is local validation, because this library carries no JSON Schema validator: the response comes back unvalidated, there is no retry-with-feedback when a model returns the wrong shape, validationAttempts is always 1, and T is yours to assert. Prefer schema unless you genuinely hold a JSON Schema already. streamStructured is unchanged and still requires Zod.

Reading next

• Tool-use security guide — runAgent code patterns, the destructive / requiresConfirmation / maxOutputBytes flags, the approval-gate wrapper
• Content blocks reference — tool_use and tool_result block shapes
• Anthropic prompt caching docs — how cache reads are billed
• Pricing source — verify the bundled table is current