For the complete documentation index, see llms.txt. Markdown versions of all docs pages are available by appending .md to any docs URL.

Configure Amazon Bedrock as an LLM provider for agentgateway.

Configure Amazon Bedrock as an LLM provider in agentgateway.

Before you begin

1. Set up an agentgateway proxy.
2. Make sure that your Amazon credentials have access to the Bedrock models that you want to use. You can alternatively use an AWS Bedrock API key.
3. Optional: You can configure AWS IAM Identity Center to allow single sign-on (SSO) credentials to authenticate to AWS Bedrock. Make sure that you have access to AWS Bedrock and set up your AWS profile to use SSO, such as through the aws CLI. Make sure the workload can use that profile (for example with AWS_PROFILE). Later when you create the AgentgatewayBackend, omit policies.auth so the proxy uses implicit AWS SSO credentials.

Set up access to Amazon Bedrock

1. Store your credentials to access the AWS Bedrock API.

   AWS credentialsAWS Bedrock API key

   Warning

   These steps store the gateway’s credentials in AGW_-prefixed environment variables, as opposed to the standard AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and AWS_SESSION_TOKEN variables that the aws CLI and the AWS SDKs read. This way, when you export the variables, you do not replace your own personal user identity for the rest of the shell session, which could lead to unexpected auth errors. To use the standard names anyway, set them in a subshell, or store the gateway’s credentials in a named AWS profile and select that profile with AWS_PROFILE for the agentgateway process only.

   1. Log in to the AWS console and store your access credentials as environment variables.

      export AGW_AWS_ACCESS_KEY_ID="<aws-access-key-id>"
      export AGW_AWS_SECRET_ACCESS_KEY="<aws-secret-access-key>"
      export AGW_AWS_SESSION_TOKEN="<aws-session-token>"

   2. Create a secret with your Bedrock API key. Optionally provide the session token.

      kubectl create secret generic bedrock-secret \
        -n agentgateway-system \
        --from-literal=accessKey="$AGW_AWS_ACCESS_KEY_ID" \
        --from-literal=secretKey="$AGW_AWS_SECRET_ACCESS_KEY" \
        --from-literal=sessionToken="$AGW_AWS_SESSION_TOKEN" \
        --type=Opaque \
        --dry-run=client -o yaml | kubectl apply -f -

   1. Save the API key in an environment variable.

      export BEDROCK_API_KEY=<insert your API key>

   2. Create a Kubernetes secret to store your Amazon Bedrock API key.

      kubectl apply -f- <<EOF
      apiVersion: v1
      kind: Secret
      metadata:
        name: bedrock-secret
        namespace: agentgateway-system
      type: Opaque
      stringData:
        Authorization: $BEDROCK_API_KEY
      EOF

2. Create an AgentgatewayBackend resource to configure your LLM provider. Make sure to reference the secret that holds your credentials to access the LLM.

   kubectl apply -f- <<EOF
   apiVersion: agentgateway.dev/v1alpha1
   kind: AgentgatewayBackend
   metadata:
     name: bedrock
     namespace: agentgateway-system
   spec:
     ai:
       provider:
         bedrock:
           model: "amazon.nova-micro-v1:0"
           region: "us-east-1"
     policies:
       auth:
         aws:
           secretRef:
             name: bedrock-secret
   EOF

   Review the following table to understand this configuration. For more information, see the API reference.

   Setting	Description
   ai.provider.bedrock	Define the LLM provider that you want to use. The example uses Amazon Bedrock.
   bedrock.model	The model to use to generate responses. In this example, you use the amazon.nova-micro-v1:0 model. Keep in mind that some models support cross-region inference. These models begin with a us. prefix, such as us.anthropic.claude-sonnet-4-20250514-v1:0. For more models, see the AWS Bedrock docs.
   bedrock.region	The AWS region where your Bedrock model is deployed. Multiple regions are not supported.
   policies.auth	Provide the credentials to use to access the Amazon Bedrock API. The example refers to the secret that you previously created. To use implicit credentials from the workload or environment instead (for example IRSA and AWS IAM Identity Center (SSO) profiles), omit the auth settings.

3. Create an HTTPRoute resource to route requests through your agentgateway proxy to the Bedrock AgentgatewayBackend.

   OpenAI-compatible v1/chat/completionsCustom route

   kubectl apply -f- <<EOF
   apiVersion: gateway.networking.k8s.io/v1
   kind: HTTPRoute
   metadata:
     name: bedrock
     namespace: agentgateway-system
   spec:
     parentRefs:
       - name: agentgateway-proxy
         namespace: agentgateway-system
     rules:
     - matches:
       - path:
           type: PathPrefix
           value: /v1/chat/completions
       backendRefs:
       - name: bedrock
         namespace: agentgateway-system
         group: agentgateway.dev
         kind: AgentgatewayBackend
   EOF

   kubectl apply -f- <<EOF
   apiVersion: gateway.networking.k8s.io/v1
   kind: HTTPRoute
   metadata:
     name: bedrock
     namespace: agentgateway-system
   spec:
     parentRefs:
       - name: agentgateway-proxy
         namespace: agentgateway-system
     rules:
     - matches:
       - path:
           type: PathPrefix
           value: /bedrock
       backendRefs:
       - name: bedrock
         namespace: agentgateway-system
         group: agentgateway.dev
         kind: AgentgatewayBackend
   EOF

4. Send a request to the LLM provider API along the route that you previously created, such as /bedrock or /v1/chat/completions depending on your route configuration. The request body must be in OpenAI chat-completions format. Verify that the request succeeds and that you get back a response from the chat completion API.

   OpenAI-compatible v1/chat/completionsCustom route

   Cloud Provider LoadBalancer:

   curl "$INGRESS_GW_ADDRESS/v1/chat/completions" -H content-type:application/json -d '{
       "model": "",
       "messages": [
         {
           "role": "user",
           "content": "You are a cloud native solutions architect, skilled in explaining complex technical concepts such as API Gateway, microservices, LLM operations, kubernetes, and advanced networking patterns. Write me a 20-word pitch on why I should use an AI gateway in my Kubernetes cluster."
         }
       ]
     }' | jq

   Localhost:

   curl "localhost:8080/v1/chat/completions" -H content-type:application/json -d '{
       "model": "",
       "messages": [
         {
           "role": "user",
           "content": "You are a cloud native solutions architect, skilled in explaining complex technical concepts such as API Gateway, microservices, LLM operations, kubernetes, and advanced networking patterns. Write me a 20-word pitch on why I should use an AI gateway in my Kubernetes cluster."
         }
       ]
     }' | jq

   Cloud Provider LoadBalancer:

   curl "$INGRESS_GW_ADDRESS/bedrock" -H content-type:application/json -d '{
       "model": "",
       "messages": [
         {
           "role": "user",
           "content": "You are a cloud native solutions architect, skilled in explaining complex technical concepts such as API Gateway, microservices, LLM operations, kubernetes, and advanced networking patterns. Write me a 20-word pitch on why I should use an AI gateway in my Kubernetes cluster."
         }
       ]
     }' | jq

   Localhost:

   curl "localhost:8080/bedrock" -H content-type:application/json -d '{
       "model": "",
       "messages": [
         {
           "role": "user",
           "content": "You are a cloud native solutions architect, skilled in explaining complex technical concepts such as API Gateway, microservices, LLM operations, kubernetes, and advanced networking patterns. Write me a 20-word pitch on why I should use an AI gateway in my Kubernetes cluster."
         }
       ]
     }' | jq

   Example output. Note that agentgateway returns OpenAI-shaped responses, including OpenAI-style usage fields (prompt_tokens, completion_tokens, total_tokens), even though the upstream provider is Bedrock.

   {
     "id": "chatcmpl-abc123",
     "object": "chat.completion",
     "created": 1730000000,
     "model": "amazon.nova-micro-v1:0",
     "choices": [
       {
         "index": 0,
         "message": {
           "role": "assistant",
           "content": "An AI gateway in your Kubernetes cluster can enhance performance, scalability, and security while simplifying complex operations. It provides a centralized entry point for AI workloads, automates deployment and management, and ensures high availability."
         },
         "finish_reason": "stop"
       }
     ],
     "usage": {
       "prompt_tokens": 60,
       "completion_tokens": 47,
       "total_tokens": 107
     }
   }

Bedrock Mantle

Bedrock serves models on two API surfaces: the Runtime endpoint, which carries the Converse and Invoke APIs, and the Mantle endpoint, which carries the native OpenAI and Anthropic APIs. Some models are served on only one of the two.

For chat requests, the endpoint is chosen per model from the runtime and mantle tags in your model cost catalog. Run agctl catalog import to populate those tags, because the default sources include aws-bedrock-mantle, which reads them from the AWS model cards. Without a catalog, no model carries either tag, so every chat request falls back to the preference alone.

Set spec.ai.provider.bedrock.endpointPreference on the AgentgatewayBackend resource to choose how the tags are applied. The AgentgatewayModel resource takes the same setting at spec.bedrock.endpointPreference.

spec:
  ai:
    provider:
      bedrock:
        model: "amazon.nova-micro-v1:0"
        region: "us-east-1"
        endpointPreference: RuntimePreferred

Standalone mode takes the same four values in lowercase, such as runtimePreferred, under params.bedrockEndpointPreference. A value that you copy from one mode to the other fails to load.

The preference applies to chat completions, messages, responses, and Anthropic token counting. The other route types ignore it: embeddings, reranking, realtime, Gemini token counting, detection, passthrough, and content generation always take Runtime, and model listing always takes Mantle.

Whether the preference changes the request format that a model accepts depends on the endpoint that it selects. A model that resolves to Runtime accepts the Bedrock Converse format only, and its chat format tags do not apply. A model that resolves to Mantle accepts the formats in its tags, except for anthropic.claude* models, which always take the Anthropic Messages format. For more information, see Chat format tags.

Prompt caching

Prompt Caching is a performance, cost-optimization, and cost-reduction feature that allows the model to “remember” frequently used parts of your prompt, including long system instructions, reference documents, or tool definitions. This way, the model does not need to reprocess these parts every time you send a new prompt.

For example, let’s assume you have a 50-page manual and you want to ask your model different questions about the manual. Instead of re-reading the manual for each question, the model can read it once and save it in its internal cache. Then, the model can answer subsequent questions more quickly and more cost efficient.

Prompt caching is configured by using the backend.ai.promptCaching fields in the AgentgatewayPolicy resource.

1. Create an AgentgatewayPolicy resource with your prompt cache settings. The following example enables caching for system prompts and conversation messages, but disables it for tool definitions. Bedrock requires you to set the minimum token count after which caching is enabled. By default, a minimum of 1024 tokens are required by Bedrock for caching to be effective. This is also referred to as a caching checkpoint. For more information, see the API reference.

   kubectl apply -f- <<EOF
   apiVersion: agentgateway.dev/v1alpha1
   kind: AgentgatewayPolicy
   metadata:
     name: bedrock-caching-policy
     namespace: agentgateway-system
   spec:
     targetRefs:
       - group: gateway.networking.k8s.io
         kind: HTTPRoute
         name: bedrock
     backend:
       ai:
         promptCaching:
           cacheSystem: true
           cacheMessages: true
           cacheTools: false
           minTokens: 1024
   EOF

2. Port-forward the agentgateway proxy on port 15000.

   kubectl port-forward deploy/agentgateway-proxy -n agentgateway-system 15000

3. Get the caching configuration and verify that you see the cache settings.

   curl -s http://localhost:15000/config_dump | jq '.policies[] |                                    
    select(.name.name == "bedrock-caching-policy" and 
         .policy.backend.aI.promptCaching != null)'

   Example output:

   {
      "key": "backend/agentgateway-system/bedrock-caching-policy:ai:agentgateway-system/bedrock",
      "name": {
        "kind": "AgentgatewayPolicy",
        "name": "bedrock-caching-policy",
        "namespace": "agentgateway-system"
      },
      "target": {
        "route": {
          "name": "bedrock",
          "namespace": "agentgateway-system",
          "kind": "HTTPRoute"
        }
      },
      "policy": {
        "backend": {
          "aI": {
            "defaults": {},
            "overrides": {},
            "promptCaching": {
              "cacheSystem": true,
              "cacheMessages": true,
              "cacheTools": false,
              "minTokens": 1024
            }
          }
        }
      }
   }
   

Extended thinking and reasoning

Extended thinking and reasoning lets models reason through complex problems before generating a response. You can opt in to extended thinking and reasoning by adding the OpenAI reasoning_effort field to your request. Agentgateway translates this setting to Bedrock’s native thinking budget automatically.

Note: Extended thinking and reasoning requires a Claude model that supports it, such as us.anthropic.claude-opus-4-20250514-v1:0.

Use the reasoning_effort field to control how much reasoning the model applies. The value is automatically mapped to a thinking budget.

For Claude models that support adaptive thinking, the request is sent with thinking.type set to adaptive and the effort level in output_config.effort, instead of a thinking budget. The value minimal is sent as low. Which models take this form depends on the adaptive_thinking tag in the model cost catalog. The built-in catalog sets this tag for these models.

reasoning_effort also works with the following non-Claude model families. The family is chosen by matching the model ID, and the value is added to the additionalModelRequestFields of the Bedrock request.

Any other model ID is treated as a Claude model.

Cloud Provider LoadBalancer:

curl "$INGRESS_GW_ADDRESS/v1/chat/completions" -H content-type:application/json -d '{
  "model": "",
  "max_tokens": 6000,
  "reasoning_effort": "high",
  "messages": [
    {
      "role": "user",
      "content": "Explain the trade-offs between consistency and availability in distributed systems."
    }
  ]
}' | jq

Localhost:

curl "localhost:8080/v1/chat/completions" -H content-type:application/json -d '{
  "model": "",
  "max_tokens": 6000,
  "reasoning_effort": "high",
  "messages": [
    {
      "role": "user",
      "content": "Explain the trade-offs between consistency and availability in distributed systems."
    }
  ]
}' | jq

Structured outputs

Structured outputs constrain the model to respond with a specific JSON schema. Provide the schema definition in the OpenAI response_format field of your request. Agentgateway translates this to Bedrock’s native format automatically.

Cloud Provider LoadBalancer:

curl "$INGRESS_GW_ADDRESS/v1/chat/completions" -H content-type:application/json -d '{
  "model": "",
  "max_tokens": 256,
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "answer_schema",
      "schema": {
        "type": "object",
        "properties": {
          "answer": { "type": "string" },
          "confidence": { "type": "number" }
        },
        "required": ["answer", "confidence"],
        "additionalProperties": false
      }
    }
  },
  "messages": [
    {
      "role": "user",
      "content": "Is the sky blue? Respond with your answer and a confidence score between 0 and 1."
    }
  ]
}' | jq

Localhost:

curl "localhost:8080/v1/chat/completions" -H content-type:application/json -d '{
  "model": "",
  "max_tokens": 256,
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "answer_schema",
      "schema": {
        "type": "object",
        "properties": {
          "answer": { "type": "string" },
          "confidence": { "type": "number" }
        },
        "required": ["answer", "confidence"],
        "additionalProperties": false
      }
    }
  },
  "messages": [
    {
      "role": "user",
      "content": "Is the sky blue? Respond with your answer and a confidence score between 0 and 1."
    }
  ]
}' | jq

Next steps