Keycloak Cedar Policy Engine

Why we're exploring this

Keycloak Authorization Services provides policy-based authorization, but its policy evaluation capabilities can be extended with different policy engines. This experiment explores integrating Cedar, AWS's open-source policy language, directly into Keycloak as the policy engine for authorization decisions. Policies are managed through the Keycloak Admin Console, while authorization decisions are exposed through Keycloak's AuthZEN endpoint. The result is a Keycloak-based authorization architecture that brings Cedar's policy language and evaluation model to both human and AI agent authorization.

How a decision flows

Every request, human or agent, goes through the same path. The gateway enforces; Keycloak and Cedar decide.

One policy language, humans and agents

Cedar policies are declarative and readable. The same engine evaluates a human's role-based access and an AI agent's delegated scope, with no separate code path.

permit(
  principal,
  action == Action::"read",
  resource
)
when {
  principal.roles
    .contains("document-viewer")
  && resource.type == "Document"
};

permit(
  principal,
  action == Action::"read",
  resource
)
when {
  principal.clientId ==
    "spiffe://org/agents/planner"
  && principal.scope
       .contains("docs:read")
  && "spiffe://org/agents/orchestrator"
       in principal.aud
  && resource.type == "Document"
};

What it does today

• → Cedar policies are evaluated in process by Keycloak's Authorization Services.
• → Keycloak remains the policy decision point (PDP), with decisions exposed through its AuthZEN endpoint.
• → The Cedar schema is automatically generated from Keycloak's resource server scopes and types.