本文為英文版的機器翻譯版本，如內容有任何歧義或不一致之處，概以英文版為準。

範例

下列範例顯示常見攔截器使用案例的 Python AWS Lambda 函數。

傳遞攔截器

此範例示範一個簡單的攔截器，記錄 REQUEST 攔截器的 MCP 方法，並透過不變的方式傳遞所有請求和回應：

import json
import logging

# Configure logging
logger = logging.getLogger()
logger.setLevel(logging.INFO)

def lambda_handler(event, context):
    """
    Lambda function that handles both REQUEST and RESPONSE interceptor types.

    For REQUEST interceptors: logs the MCP method and passes request through unchanged
    For RESPONSE interceptors: passes response through unchanged
    """
    # Extract the MCP data from the event
    mcp_data = event.get('mcp', {})

    # Check if this is a REQUEST or RESPONSE interceptor based on presence of gatewayResponse
    if 'gatewayResponse' in mcp_data and mcp_data['gatewayResponse'] != None:
        # This is a RESPONSE interceptor
        logger.info("Processing RESPONSE interceptor - passing through unchanged")

        # Pass through the original request and response unchanged
        response = {
          "interceptorOutputVersion": "1.0",
          "mcp": {
              "transformedGatewayResponse": {
                  "body": mcp_data.get('gatewayResponse', {}).get('body', {}),
                  "statusCode": mcp_data.get('gatewayResponse', {}).get('statusCode', 200)
              }
          }
        }
    else:
        # This is a REQUEST interceptor
        gateway_request = mcp_data.get('gatewayRequest', {})
        request_body = gateway_request.get('body', {})
        mcp_method = request_body.get('method', 'unknown')

        # Log the MCP method
        logger.info(f"Processing REQUEST interceptor - MCP method: {mcp_method}")

        # Pass through the original request unchanged
        response = {
          "interceptorOutputVersion": "1.0",
          "mcp": {
              "transformedGatewayRequest": {
                  "body": request_body,
              }
          }
        }

    return response

此 Lambda 函數可以同時設定為 REQUEST 和 RESPONSE 攔截器。當設定為 REQUEST 攔截器時，它會記錄來自傳入請求的 MCP 方法。設定為 RESPONSE 攔截器時，只會傳遞未變更的回應。這兩種攔截器類型都會在不修改的情況下傳回原始資料，使其成為「傳遞」攔截器。

使用請求攔截器自訂模型路由

對於推論目標，閘道會使用模型型路由從 model 欄位選取目標。REQUEST 攔截器可以在閘道評估路由model之前重寫，這可讓您支援虛擬模型 ID：未對應至任何單一設定模型的穩定別名。攔截器會將別名解析為格式為 的具體目標合格 ID{targetName}/{modelId}，因此呼叫者會在您集中控制模型選擇時使用一個名稱。

下列 REQUEST 攔截器會根據請求輸入的大小，將虛擬模型 ID 解析auto-claude為特定 Anthropic 模型：anthropic/claude-haiku-4-5適用於小型請求、anthropic/claude-sonnet-4-6適用於中型請求，以及anthropic/claude-opus-4-7適用於大型請求。呼叫者傳送 auto-claude，而閘道會將每個請求路由到anthropic目標上已解析的模型。使用任何其他模型 ID 的請求會傳遞不變，並遵循正常的模型型路由。

import base64
import json
import logging

logger = logging.getLogger()
logger.setLevel(logging.INFO)

# The virtual model ID that this interceptor resolves to a concrete
# target-qualified ID in the form "{targetName}/{modelId}".
VIRTUAL_MODEL = "auto-claude"

# Approximate input-size thresholds (in characters) used to choose a model.
SONNET_THRESHOLD = 2000
OPUS_THRESHOLD = 8000


def resolve_auto_claude(payload):
    """Choose an Anthropic model based on the size of the request input."""
    input_size = len(json.dumps(payload.get("input", "")))
    if input_size >= OPUS_THRESHOLD:
        return "anthropic/claude-opus-4-7"
    if input_size >= SONNET_THRESHOLD:
        return "anthropic/claude-sonnet-4-6"
    return "anthropic/claude-haiku-4-5"


def lambda_handler(event, context):
    http = event.get("http", {})
    request = http.get("gatewayRequest", {})
    encoded_body = request.get("body")

    # Nothing to transform (for example, an empty body) - pass through.
    if not encoded_body:
        return {"interceptorOutputVersion": "1.0", "http": {}}

    try:
        payload = json.loads(base64.b64decode(encoded_body))
    except (ValueError, TypeError) as exc:
        logger.warning("Passing request through unchanged; could not parse body: %s", exc)
        return {"interceptorOutputVersion": "1.0", "http": {}}

    # Only resolve the virtual model ID; let everything else route normally.
    if not isinstance(payload, dict) or payload.get("model") != VIRTUAL_MODEL:
        return {"interceptorOutputVersion": "1.0", "http": {}}

    resolved = resolve_auto_claude(payload)
    logger.info("Resolved virtual model %s -> %s", VIRTUAL_MODEL, resolved)
    payload["model"] = resolved

    encoded = base64.b64encode(json.dumps(payload).encode("utf-8")).decode("utf-8")
    return {
        "interceptorOutputVersion": "1.0",
        "http": {
            "transformedGatewayRequest": {
                "body": encoded
            }
        },
    }

將 函數連接至您的閘道做為 REQUEST 攔截器。如需說明，請參閱設定攔截器。閘道服務角色也必須具有叫用 函數的許可。如需詳細資訊，請參閱攔截器的許可。

若要查看解決方案，請傳送使用auto-claude別名的請求。攔截器會根據輸入大小重寫model至具體的 Anthropic 模型，而閘道會將請求路由至anthropic目標上的該模型。

awscurl --service bedrock-agentcore --region us-west-2 -X POST \
    "https://GATEWAY_ID.gateway.bedrock-agentcore.us-west-2.amazonaws.com/inference/v1/responses" \
    -H "Content-Type: application/json" \
    -d '{"model": "auto-claude", "input": "Hello!", "max_output_tokens": 50}'