Back to AI Framework Overview | All Providers

@memberjunction/ai-anthropic

MemberJunction AI provider for Anthropic's Claude models. Implements the BaseLLM interface from @memberjunction/ai, supporting the Claude 4, Claude 3.5, and Claude 3 model families including Opus, Sonnet, and Haiku variants.

Architecture

graph TD
    A["AnthropicLLM
(Provider)"] -->|extends| B["BaseLLM
(@memberjunction/ai)"]
    B -->|registered via| C["@RegisterClass
(@memberjunction/global)"]
    A -->|wraps| D["Anthropic SDK
(@anthropic-ai/sdk)"]
    A -->|provides| E["Chat Completions"]
    A -->|provides| F["Streaming"]
    A -->|provides| G["Prompt Caching"]
    A -->|provides| H["Thinking/Reasoning
Extraction"]

    style A fill:#7c5295,stroke:#563a6b,color:#fff
    style B fill:#2d6a9f,stroke:#1a4971,color:#fff
    style C fill:#b8762f,stroke:#8a5722,color:#fff
    style D fill:#2d8659,stroke:#1a5c3a,color:#fff
    style E fill:#2d6a9f,stroke:#1a4971,color:#fff
    style F fill:#2d6a9f,stroke:#1a4971,color:#fff
    style G fill:#2d6a9f,stroke:#1a4971,color:#fff
    style H fill:#2d6a9f,stroke:#1a4971,color:#fff
Copy

Features

• Chat Completions: Full support for Anthropic's Messages API
• Streaming: Real-time response streaming with thinking block extraction
• Prompt Caching: Automatic ephemeral cache control on content blocks for reduced latency and cost, with breakpoint placement that keeps a trailing volatile message (the agent runtime-state fragment) out of the cached prefix
• Multimodal Input: Support for text, images (base64 and URL), and content block arrays
• Thinking/Reasoning: Extraction of thinking content from Claude's extended thinking responses
• Error Analysis: Integrated error analysis via ErrorAnalyzer

Installation

npm install @memberjunction/ai-anthropic
Copy

Usage

import { AnthropicLLM } from "@memberjunction/ai-anthropic";

const llm = new AnthropicLLM("your-anthropic-api-key");

const result = await llm.ChatCompletion({
    model: "claude-sonnet-4-20250514",
    messages: [
        { role: "user", content: "Explain quantum computing in simple terms." },
    ],
    temperature: 0.7,
    maxOutputTokens: 1024,
});

console.log(result.data.choices[0].message.content);
Copy

Streaming

const result = await llm.ChatCompletion({
    model: "claude-sonnet-4-20250514",
    messages: [{ role: "user", content: "Write a short story." }],
    streaming: true,
    streamingCallbacks: {
        OnContent: (content) => process.stdout.write(content),
        OnComplete: (result) => console.log("\nDone!"),
    },
});
Copy

Extended Thinking

const result = await llm.ChatCompletion({
    model: "claude-opus-4-20250514",
    messages: [{ role: "user", content: "Solve this step by step: ..." }],
    effortLevel: "80",
});

// Access thinking content alongside the response
console.log("Thinking:", result.data.choices[0].message.thinking);
console.log("Answer:", result.data.choices[0].message.content);
Copy

Prompt caching and the trailing runtime-state fragment

Anthropic caches only up to an explicit cache_control breakpoint, and a read-hit requires the new request to match a cached prefix at a breakpoint. By default the driver places the breakpoint on the last message. When the last message is the agent framework's volatile runtime-state fragment — recognised by the volatileState metadata flag via BaseLLM.isVolatileStateMessage, or by a leading <mj-runtime-state> / <mj-agent-specialization> literal for callers passing plain messages — the driver uses BaseLLM.splitTrailingVolatileState to put the breakpoint on the last real history message and sends the fragment (and any assistant prefill after it) uncached. If that leaves two user turns adjacent it inserts an OK assistant turn to preserve role alternation. Result: the stable history caches across loop iterations and only the fragment is re-processed. A <<<MJ_CACHE_BREAKPOINT>>> marker inside a text block still works for callers that want an explicit boundary elsewhere. See the Agent Prompt Caching Guide.

Supported Parameters

Class Registration

Registered as AnthropicLLM via @RegisterClass(BaseLLM, 'AnthropicLLM') for use with MemberJunction's class factory system.

Dependencies

• @memberjunction/ai - Core AI abstractions
• @memberjunction/global - Class registration
• @anthropic-ai/sdk - Official Anthropic SDK