AWS launches Claude apps gateway for enterprise governance controls

Anthropic and AWS have released a production-ready reference deployment of Claude apps gateway, a self-hosted governance layer designed to manage Claude Code and Claude Desktop deployments across enterprise teams. The gateway runs in server mode, loaded via the same Claude Code CLI binary that developers already use, and can be deployed on AWS Fargate, Amazon Elastic Kubernetes Service (EKS), or Amazon Elastic Compute Cloud (EC2).

The reference architecture uses an internal Application Load Balancer to terminate TLS traffic, with Amazon Route 53 private hosted zones resolving the gateway to private IP addresses reachable through VPN or AWS Direct Connect. Session and spend-tracking state is stored in Amazon RDS for PostgreSQL, allowing any stateless gateway container to serve any request. The gateway authenticates to Amazon Bedrock using the IAM role assigned to the Fargate task, while the Claude Platform on AWS API key and other static credentials remain in AWS Secrets Manager—no upstream credentials are distributed to developer machines.

Authentication follows OAuth 2.0 device authorization grant. When a developer runs /login, the client opens a browser to authenticate through the organization's OIDC identity provider (such as Okta, Microsoft Entra ID, Auth0, Keycloak, or Amazon Cognito). The gateway issues a short-lived bearer token valid for one hour by default, with silent refresh handled in the background. This approach enables centralized single sign-on with instant offboarding—removing a user from the identity provider causes their session to expire within the TTL, with no credential rotation required. The gateway picks up group membership from the identity provider 1:1, with no translation layer or pre-created user directory.

For every inference request, the gateway validates the bearer token, resolves the developer's identity and group membership, applies the matching policy, evaluates the spend cap, and routes the request to Amazon Bedrock or Claude Platform on AWS. Policies are defined in YAML and evaluated in declaration order; the first match is selected and merged onto a catch-all base policy. For example, contractors can be restricted to Claude Haiku and Claude Sonnet 5 with web access denied, while engineers receive access to Claude Opus 4.8, Sonnet 5, and Haiku with allowed file operations and restrictions on reading secrets. Model access is enforced server-side, so a developer whose policy grants only Haiku cannot bypass this restriction even with a modified client. Changes to policies propagate to connected clients within one hour, requiring no action from developers.

The gateway relays client usage metrics (token usage, cost usage, and active time) over OpenTelemetry Protocol (OTLP) to a collector you configure, with metrics attributed to the authenticated developer's user ID, email, and group membership. Supported OTLP-compatible backends include Datadog, Splunk, Grafana, and Amazon CloudWatch through the AWS Distro for OpenTelemetry (ADOT) collector. Most deployments start with metrics only, which provide per-user cost and usage breakdowns without exposing sensitive data like source code or prompt content.

The gateway supports automatic failover across multiple upstreams in declared order, failing over on upstream unavailability, throttling, or timeouts. Organizations can configure multiple Amazon Bedrock regions or cross-provider fallback to Claude Platform on AWS. Spend caps are set at three levels—organization-wide defaults, per-group, and per-user overrides—each applying individually per developer, not as a shared pool. AWS Budgets and AWS Cost Explorer provide account-level cost visibility, while the gateway complements these tools with inline enforcement before inference occurs. Deployment scripts, configuration templates, and operational guidance are available in the accompanying repository and deployment guide.