Authorization Example
This example shows how to use the agentgateway to perform MCP native authorization. It is recommended to complete the basic example before this one.
Running the example
cargo run -- -f examples/authorization/config.yaml
In addition to the basic configuration from the basic example, we have a few new fields:
The jwtAuth indicates how to authenticate clients.
This uses example JWT keys and tokens included for demonstration purposes only.
policies:
jwtAuth:
issuer: agentgateway.dev
audiences: [test.agentgateway.dev]
jwks:
# Relative to the folder the binary runs from, not the config file
file: ./manifests/jwt/pub-key
With this configuration, users will be required to pass a valid JWT token matching the criteria.
An example token signed by the key above can be found at manifests/jwt/pub-key/example*.key; this can be
passed into the MCP inspector Authentication > Bearer Token field.
Next, we add mcpAuthorization rules. These allow specifying criteria for who is allowed to access which resources.
mcpAuthorization:
rules:
# Allow anyone to call 'echo'
- 'permit(principal, action == Action::"call_tool", resource == Tool::"echo");'
# Only the test-user can call 'add'
- 'permit(principal == User::"test-user", action == Action::"call_tool", resource == Tool::"add");'
# Any authenticated user with the claim `nested.key == value` can access 'printEnv'
- |
permit(principal is User, action == Action::"call_tool", resource == Tool::"printEnv") when
{ context.claims.nested.key == "value" };
Agentgateway authorization rules are built on CEL Policy, a small language designed for authorization policies. This allows complete flexibility to define simple or advanced access rules.
First, use the example2.key.
This key is for the user (sub) test-user and has no claims.
When we "List Tools", we can see the tools are filtered down to only echo and add, the two tools we have access to.
Next, use the example1.key.
This key is for the user (sub) test-user and has the following claims:
{
"field1": "value1",
"nested": {
"key": "value"
},
"list": ["apple", "banana"]
}
When we "List Tools", we can see the tools now additionally includes the printEnv.
This is because our third policy allows access when { context.claims.nested.key == "value" }, which now matches.
Tip
When reconnecting, MCP inspector will not clear the cached tools list. Ensure you re-run "List Tools".
When the policy doesn't allow the connecting user to access a tool, it is automatically filtered from the tools list. If the user does still attempt to call the tool, it is also denied.


