CEL-based RBAC

Use Common Language Expressions (CEL) expressions to secure access to AI resources.

About CEL-based RBAC

Agentgateway proxies use CEL expressions to match requests or responses on specific parameters, such as a request header or source address. If the request matches the condition, it is allowed. Requests that do not match any of the conditions are denied.

For an overview of supported CEL expressions, see the agentgateway docs.

Before you begin

Set up an agentgateway proxy.

Set up access to Gemini

Configure access to an LLM provider such as Gemini. You can use any other LLM provider, an MCP server, or an agent to try out CEL-based RBAC.

Save your Gemini API key as an environment variable. To retrieve your API key, log in to the Google AI Studio and select API Keys.
```
export GOOGLE_KEY=<your-api-key>
```

Create a secret to authenticate to Google.

kubectl apply -f - <<EOF
apiVersion: v1
kind: Secret
metadata:
  name: google-secret
  namespace: kgateway-system
type: Opaque
stringData:
  Authorization: $GOOGLE_KEY
EOF

Create an AgentgatewayBackend resource to configure an LLM provider that references the AI API key secret.

kubectl apply -f- <<EOF
apiVersion: agentgateway.dev/v1alpha1
kind: AgentgatewayBackend
metadata:
  name: google
  namespace: kgateway-system
spec:
  ai:
    provider:
      gemini:
        model: gemini-2.5-flash-lite
  policies:
    auth:
      secretRef:
        name: google-secret
EOF

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

Setting	Description
`ai.provider.gemini`	Define the Gemini provider.
`gemini.model`	The model to use to generate responses. In this example, you use the `gemini-2.5-flash-lite` model. For more models, see the Google AI docs.
`policies.auth`	The authentication token to use to authenticate to the LLM provider. The example refers to the secret that you created in the previous step.

Create an HTTPRoute resource that routes incoming traffic to the AgentgatewayBackend. The following example sets up a route on the /openai path to the AgentgatewayBackend that you previously created. The URLRewrite filter rewrites the path from /openai to the path of the API in the LLM provider that you want to use, /v1/chat/completions.

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

Send a request to the LLM provider API. Verify that the request succeeds and that you get back a response from the chat completion API.

curl -vik "$INGRESS_GW_ADDRESS/gemini" -H content-type:application/json  -d '{
  "model": "",
  "messages": [
   {"role": "user", "content": "Explain how AI works in simple terms."}
 ]
}'

curl -vik "localhost:8080/gemini" -H content-type:application/json  -d '{
  "model": "",
  "messages": [
   {"role": "user", "content": "Explain how AI works in simple terms."}
 ]
}'

Example output:

{"id":"aGLEaMjbLp6p_uMPopeAoAc",
"choices":
  [{"index":0,"message":{
      "content":"Imagine teaching a dog a trick.  You show it what to do, reward it when it's right, and correct it when it's wrong.  Eventually, the dog learns.\n\nAI is similar.  We \"teach\" computers by showing them lots of examples.  For example, to recognize cats in pictures, we show it thousands of pictures of cats, labeling each one \"cat.\"  The AI learns patterns in these pictures – things like pointy ears, whiskers, and furry bodies – and eventually, it can identify a cat in a new picture it's never seen before.\n\nThis learning process uses math and algorithms (like a secret code of instructions) to find patterns and make predictions.  Some AI is more like a dog learning tricks (learning from examples), and some is more like following a very detailed recipe (following pre-programmed rules).\n\nSo, in short: AI is about teaching computers to learn from data and make decisions or predictions, just like we teach dogs tricks.\n",
      "role":"assistant"
      },
   "finish_reason":"stop"
   }],
 "created":1757700714,
 "model":"gemini-1.5-flash-latest",
 "object":"chat.completion",
 "usage":{
     "prompt_tokens":8,
     "completion_tokens":205,
     "total_tokens":213
     }
}

Set up RBAC permissions

Create a AgentgatewayPolicy with your CEL rules. The following example allows requests with the x-llm: gemini header.

kubectl apply -f- <<EOF
apiVersion: agentgateway.dev/v1alpha1
kind: AgentgatewayPolicy
metadata:
  name: rbac-policy
  namespace: kgateway-system
spec:
  targetRefs:
    - group: gateway.networking.k8s.io
      kind: HTTPRoute
      name: google
  traffic: 
    authorization: 
      action: Allow
      policy: 
        matchExpressions:
          - "request.headers['x-llm'] == 'gemini'"
EOF

Send a request to the LLM provider API without the llm header. Verify that the request is denied with a 403 HTTP response code.

curl -vik "$INGRESS_GW_ADDRESS:8080/gemini" -H content-type:application/json  -d '{
  "model": "",
  "messages": [
   {"role": "user", "content": "Explain how AI works in simple terms."}
 ]
}'

curl -vik "localhost:8080/gemini" -H content-type:application/json  -d '{
  "model": "",
  "messages": [
   {"role": "user", "content": "Explain how AI works in simple terms."}
 ]
}'

Example output:

* upload completely sent off: 109 bytes
< HTTP/1.1 403 Forbidden
HTTP/1.1 403 Forbidden
< content-type: text/plain
content-type: text/plain
< content-length: 20
content-length: 20
< 

* Connection #0 to host localhost left intact
authorization failed

Send another request to the LLM provider. This time, you include the llm header. Verify that the request succeeds with a 200 HTTP response code.

curl -vik "$INGRESS_GW_ADDRESS:8080/gemini" \
  -H "content-type: application/json" \
  -H "x-llm: gemini" -d '{
  "model": "",
  "messages": [
   {"role": "user", "content": "Explain how AI works in simple terms."}
 ]
}'

curl -vik "localhost:8080/gemini" \
  -H "content-type: application/json" \
  -H "x-llm: gemini" -d '{
  "model": "",
  "messages": [
   {"role": "user", "content": "Explain how AI works in simple terms."}
 ]
}'

Cleanup

You can remove the resources that you created in this guide.

kubectl delete AgentgatewayPolicy rbac-policy -n kgateway-system
kubectl delete httproute google -n kgateway-system
kubectl delete AgentgatewayBackend google -n kgateway-system
kubectl delete secret google-secret -n kgateway-system

Inference routing Configuration