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

Send requests through agentgateway in the native Gemini wire format, including streaming and token counting.

Serve clients that send the native Gemini wire format, such as the Gemini and Vertex AI SDKs, through agentgateway.

About

The Gemini API addresses a model through the request path rather than the request body, as models/{model}:generateContent. A client that is built on the Gemini or Vertex AI SDK sends this format directly, so pointing that client at agentgateway needs no OpenAI-compatible shim.

Two route types carry the format:

Because the model name comes from the path, any Gemini model works without a per-model entry in your configuration. The tunedModels/{model} form is preserved as well.

Native Gemini requests reach Gemini-family backends only:

• The Gemini API
• Vertex AI with a Gemini model

A native Gemini request that is routed to any other provider is rejected with an unsupported-conversion error rather than translated.

Before you begin

1. Install and set up an agentgateway proxy.
2. Set up access to the Gemini or Vertex AI LLM provider.

Step 1: Add the Gemini route types

Create an AgentgatewayBackend that maps the Gemini method suffixes to their route types. The default behavior routes all traffic as Completions, so the native Gemini paths must be mapped explicitly.

If you already set up multiple endpoints, add these paths to your existing AgentgatewayBackend.

kubectl apply -f- <<EOF
apiVersion: agentgateway.dev/v1alpha1
kind: AgentgatewayBackend
metadata:
  name: google-native
  namespace: agentgateway-system
spec:
  ai:
    provider:
      gemini: {}
  policies:
    auth:
      secretRef:
        name: google-secret
    ai:
      routes:
        ":generateContent": "GenerateContent"
        ":streamGenerateContent": "GenerateContent"
        ":countTokens": "GeminiCountTokens"
EOF

Review the following table to understand this configuration.

Step 2: Create an HTTPRoute

Create an HTTPRoute that routes the native Gemini paths to the AgentgatewayBackend.

kubectl apply -f- <<EOF
apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
  name: google-native
  namespace: agentgateway-system
spec:
  parentRefs:
  - name: agentgateway-proxy
    namespace: agentgateway-system
  rules:
  - matches:
    - path:
        type: PathPrefix
        value: /v1beta/models
    backendRefs:
    - name: google-native
      namespace: agentgateway-system
      group: agentgateway.dev
      kind: AgentgatewayBackend
EOF

Step 3: Send a native Gemini request

Send a request to models/{model} with the Gemini method suffix. Agentgateway forwards the body unchanged and returns the Gemini response shape to the client.

curl "$INGRESS_GW_ADDRESS/v1beta/models/gemini-2.5-flash:generateContent" \
  -H content-type:application/json \
  -d '{
    "contents": [{"role": "user", "parts": [{"text": "Say hello"}]}]
  }' | jq

curl "localhost:8080/v1beta/models/gemini-2.5-flash:generateContent" \
  -H content-type:application/json \
  -d '{
    "contents": [{"role": "user", "parts": [{"text": "Say hello"}]}]
  }' | jq

To estimate the size of a request before you send it, use the :countTokens suffix with the same body.

Step 4: Stream a response

Streaming uses the :streamGenerateContent suffix, and requires the alt=sse query parameter.

curl -N "$INGRESS_GW_ADDRESS/v1beta/models/gemini-2.5-flash:streamGenerateContent?alt=sse" \
  -H content-type:application/json \
  -d '{
    "contents": [{"role": "user", "parts": [{"text": "Count to five"}]}]
  }'

curl -N "localhost:8080/v1beta/models/gemini-2.5-flash:streamGenerateContent?alt=sse" \
  -H content-type:application/json \
  -d '{
    "contents": [{"role": "user", "parts": [{"text": "Count to five"}]}]
  }'

Without alt=sse, the Gemini API streams a JSON array instead of server-sent events, which agentgateway cannot parse incrementally. Rather than fail partway through a response, the request is rejected before it reaches the provider.

{
  "error": {
    "code": 400,
    "message": "streamGenerateContent requires alt=sse; the JSON-array streaming variant is not supported",
    "status": "INVALID_ARGUMENT"
  }
}

Cleanup

kubectl delete httproute google-native -n agentgateway-system
kubectl delete AgentgatewayBackend google-native -n agentgateway-system