product updateAnthropic

AWS Publishes Reference Architecture for Deploying Anthropic's Claude Apps Gateway at Enterprise Scale

TL;DR

AWS published a production reference architecture for deploying Anthropic's Claude apps gateway, a self-hosted governance layer that sits between Claude Code, Claude Desktop, and Amazon Bedrock or Claude Platform on AWS. The deployment pattern centralizes SSO authentication, model access policy, and spend controls for enterprise rollouts.

3 min read
0

AWS has published a production reference deployment for Anthropic's Claude apps gateway, a self-hosted governance layer designed to sit between Claude Code, Claude Desktop, and either Amazon Bedrock or Claude Platform on AWS. The post extends an earlier launch announcement with concrete architecture, deployment patterns, and configuration guidance for enterprise administrators.

The gateway ships inside the same Claude Code CLI binary developers already use and starts in server mode via claude gateway --config gateway.yaml. In AWS's reference architecture, it runs as a stateless container on AWS Fargate within a customer VPC, though AWS notes the same image also runs on Amazon EKS or Amazon EC2.

Architecture details. Session state — including device codes, sign-in sessions, per-user spend counters, and audit records — lives in Amazon RDS for PostgreSQL rather than in the gateway task itself. Because auth state is externalized, any Fargate task can serve any request without sticky sessions on the load balancer. An internal Application Load Balancer terminates TLS via AWS Certificate Manager, with Amazon Route 53 private hosted zones resolving the gateway over VPN or AWS Direct Connect. The gateway authenticates to Bedrock using its assigned IAM role, while Claude Platform API keys and other static credentials stay in AWS Secrets Manager — no upstream credentials are distributed to developer machines.

Governance controls. The gateway addresses five stated governance needs, with the post detailing two in depth:

  • Identity/SSO: Authentication delegates entirely to a customer's OIDC provider (Okta, Microsoft Entra ID, Auth0, Keycloak, or Amazon Cognito). Developers sign in once via browser-based OAuth 2.0 device authorization grant; the gateway issues a bearer token with a one-hour default time-to-live and refreshes it silently. The gateway maintains no user directory of its own — group membership comes directly from the identity provider, and offboarding is accomplished simply by removing a user from that provider.

  • Policy/model access: Administrators define model access and tool permissions in a single YAML policy block, matched by identity provider group and evaluated in declaration order with a required catch-all (match: {}) entry. Example policies shown restrict contractor groups to Claude Haiku models with web access denied, while engineering groups get access to Claude Opus, Sonnet, and Haiku models with file-access guardrails. Policy changes propagate to connected clients within one hour with no developer action required.

Usage metrics are attributed to authenticated identities and forwarded via the OpenTelemetry Protocol (OTLP) to a customer-configured collector, enabling per-user or per-group cost attribution and spend-cap enforcement.

AWS did not disclose pricing for running the gateway infrastructure itself (Fargate, RDS, ALB, and NAT gateway costs apply per standard AWS billing), nor pricing changes to underlying Bedrock or Claude Platform inference. Deployment scripts, configuration templates, and an operational guide are available through AWS's accompanying repository and documentation.

What this means: This is an infrastructure and governance release, not a new model. It targets a specific enterprise pain point — IT and security teams need centralized SSO, model-access policy, and cost attribution before they'll approve broad rollout of Claude Code and Claude Desktop across a workforce. By keeping credentials off developer machines and externalizing session state to RDS, AWS and Anthropic are positioning the gateway as infrastructure that satisfies enterprise compliance and audit requirements rather than adding new model capability. Organizations already standardized on AWS and an OIDC identity provider get a relatively low-friction path to deploy Claude tooling at scale; the more interesting signal is that AI coding tools are now mature enough to need this kind of governance tooling in the first place.

Related Articles

product update

First Orion Cuts QA Bottlenecks by Replacing Selenium Scripts with Amazon Nova Act Agents

Branded communications company First Orion adopted Amazon Nova Act as a pre-release partner in March 2025 to replace fragile Selenium and Playwright test scripts with natural-language QA automation. The shift let QA analysts author tests directly without waiting on automation engineers to translate test cases into code.

product update

Anthropic Makes Claude Code's Auto Mode Default for Pro, Max, and Team Users on August 14

Anthropic will make Claude Code's auto mode the default for Pro, Max, and Team accounts starting August 14, reducing step-by-step approval prompts. The company cites a study of 1,053 testers showing auto mode caught 89% of harmful actions versus 13.6% for manual review.

product update

Anthropic Makes Auto Mode Default in Claude Code for Pro, Max, and Team Plans Starting August 14

Anthropic will make auto mode the default setting for new Claude Code sessions on Pro, Max, and Team plans starting August 14, 2026. The company cites a 1,053-person study showing auto mode blocked 89% of harmful actions compared to 13.6% for human reviewers, plus a third-party test claiming zero successful prompt injections out of 720 attempts.

product update

Anthropic Makes Claude Code's Auto Mode Default, Cutting Dangerous Command Approvals by 75 Percentage Points

Anthropic will enable Auto Mode by default in Claude Code for Pro, Max, and Team plans starting August 14, 2026. Internal testing with 1,053 paid users found the classifier-based system caught 89% of dangerous commands compared to 13.6% caught by human reviewers, while producing 25% more pull requests.

Comments

Loading...