purra-anthropic · Python

English | 简体中文

Anthropic Messages gateway for PurrA using the official Python SDK. Supports text, function tools, streaming, cancellation, and token usage. Requires Python 3.11+.

Install

From the repository root:

python -m pip install . ./integrations/anthropic/python

Configure

from anthropic import AsyncAnthropic
from purra_anthropic import AnthropicMessagesGateway

client = AsyncAnthropic()
gateway = AnthropicMessagesGateway(client, timeout_seconds=60)

Pass gateway to AgentCore(model_gateway=gateway, ...). In the Run's ModelRequest, set provider="anthropic", the model ID, and its capability snapshot. The application supplies credentials and verified model limits.

Model options

ModelRequest.options accepts max_tokens, temperature, top_p, top_k, thinking, and output_config. Choose options supported by the selected model.

Enabled reasoning requires an explicit thinking configuration, such as {"type": "adaptive"} on a supporting model. A manual budget_tokens must be at least 1024 and below the resolved output limit. Forced tool choice cannot be combined with enabled thinking.

Behavior

SDK retries are disabled; Core controls retry and budget accounting. The gateway applies Core's resolved output limit and combines system/developer instructions into Anthropic's system field.

Signed thinking and redacted blocks are retained as private continuation data. Keep them with the unchanged assistant message when continuing on the same model; they are excluded from public output. Input usage includes cache writes and reads.

Multimodal content, hosted tools, and beta APIs are not supported. Other vendors require application adapters. Close a supplied SDK client in the application; await gateway.close() only closes a client created by the gateway.