LangChain JS (ChatAnthropic)

ChatAnthropic from @langchain/anthropic, on plugboard's Messages API. For the OpenAI-shaped client, see LangChain JS.

Connect

anthropicApiUrl is the plugboard origin, without /v1:

export function chatModel(model: string, overrides: Record<string, unknown> = {}) {
  return new ChatAnthropic({
    model,
    anthropicApiUrl: plugboardURL,
    anthropicApiKey: plugboardKey,
    maxTokens: 1024,
    maxRetries: 0,
    ...overrides,
  })
}

Tool calls

const llm = chatModel(model).bindTools([getWeather])
const first = await llm.invoke([new HumanMessage(PROMPT)])
const call = first.tool_calls?.[0]
if (!call) throw new Error("turn 1 had no tool call")
const second = await llm.invoke([new HumanMessage(PROMPT), first, new ToolMessage({ tool_call_id: call.id!, content: JSON.stringify(weatherResult) })])

Streaming

for await (const chunk of await chatModel(model).stream(PROMPT)) {
  text += typeof chunk.content === "string" ? chunk.content : chunk.content.map((part) => ("text" in part ? part.text : "")).join("")
  // The stop reason arrives on the message_delta chunk, in additional_kwargs (not response_metadata).
  if (typeof chunk.additional_kwargs.stop_reason === "string") finish = finishReason(chunk.additional_kwargs.stop_reason)
  if (chunk.usage_metadata) usage = chunk.usage_metadata
}

The stop reason of a streamed answer is on the last chunk, in additional_kwargs.stop_reason.

Structured output

const output = await chatModel(model).withStructuredOutput(Meal, { name: "record_meal" }).invoke(PROMPT)

Limits

• ChatAnthropic has no model-list call. GET /v1/models with your key works.
• Errors carry error.type and no error.code.