Anthropic

Anthropic Adapter

AgentRuntime implementation backed by a native client that posts directly to the Anthropic Messages API (https://api.anthropic.com/v1/messages) over java.net.http.HttpClient — no SDK dependency, no OpenAI-compatible translation layer in between. Streaming text_delta events are forwarded straight to the StreamingSession, and Anthropic’s tool_use → tool_result loop drives @AiTool calls.

Maven Coordinates

<dependency>
    <groupId>org.atmosphere</groupId>
    <artifactId>atmosphere-anthropic</artifactId>
    <version>${project.version}</version>
</dependency>

How It Works

Drop the dependency alongside atmosphere-ai and the framework auto-detects it via ServiceLoader. AnthropicAgentRuntime has priority 100, taking precedence over the built-in client (priority 0):

@AiEndpoint(path = "/ai/chat", systemPrompt = "You are a helpful assistant")
public class MyChat {

    @Prompt
    public void onPrompt(String message, StreamingSession session) {
        session.stream(message);  // uses the Anthropic Messages API automatically
    }
}

Configuration

The runtime reads the framework’s resolved llm.api-key, and falls back to system properties so standalone tests and CLI usage work without an Atmosphere-wide config:

Property	Purpose	Default
`anthropic.api.key`	API key (falls back from `llm.api-key`)	—
`anthropic.version`	`anthropic-version` request header	`2023-06-01`
`anthropic.base.url`	Override for a sovereign / proxy endpoint	`https://api.anthropic.com`
`anthropic.max.tokens`	`max_tokens` on each request	client default

The model is configured through the framework’s llm.model setting. The runtime stays unavailable until an API key is resolved — it never advertises availability on classpath presence alone (Runtime Truth).

Samples

Spring Boot AI Chat — drop atmosphere-anthropic on the classpath and the same @AiEndpoint code switches to Claude