AWS ships Claude Apps Gateway for Claude Code governance

Key Takeaways

• Claude Apps Gateway routes inference requests through Amazon Bedrock or Claude Platform on AWS, replacing per-developer credentials with centralized SSO
• The gateway enforces spend caps at organization, group, or user levels with daily, weekly, and monthly limits
• All three major cloud vendors (AWS, Google Cloud, Anthropic) have now published deployment guidance for the same control plane

AWS released the Claude Apps Gateway, a self-hosted control plane that gives organizations centralized governance over Claude Code and Claude Desktop. The gateway routes inference requests to Amazon Bedrock or Claude Platform on AWS, replacing the chaos of per-developer API keys, manually distributed settings, and fragmented spend tracking that has plagued enterprise AI rollouts.

The launch follows Anthropic's introduction of the gateway a week earlier for Amazon Bedrock and Google Cloud. Google Cloud published its own Cloud Run deployment walkthrough, which means all three vendors now offer first-party guidance for the same control plane. That's notable. It signals that AI developer tooling is maturing past the "every team fends for itself" phase into something resembling actual enterprise infrastructure.

What does the Claude Apps Gateway actually do?

The gateway ships inside the Claude Code CLI binary developers already have. Deployment runs as a single stateless container on Amazon ECS, EKS, or EC2 behind an internal Application Load Balancer. Amazon RDS for PostgreSQL stores sign-in state and rate-limit counters.

Five responsibilities define the gateway's scope:

• Identity: Acts as an OpenID Connect relying party. Developers sign in through browser SSO and receive short-lived tokens. Sessions expire within a configured lifetime after a user is removed from the identity provider.
• Policy: Administrators define managed settings once, scoped by identity provider group. This covers allowed models, tool permissions, and defaults developers cannot override locally.
• Telemetry: The client stamps usage metrics on every request. The gateway relays them via OpenTelemetry Protocol to Amazon CloudWatch, Amazon Managed Service for Prometheus, or your own collector.
• Routing: The gateway holds upstream credentials and forwards requests on developers' behalf, with optional failover across AWS Regions or accounts.
• Spend caps: Daily, weekly, and monthly limits per organization, group, or user. Requests stop once a cap is hit.

How does configuration work?

A single YAML file read at startup handles everything. The Bedrock upstream uses the container's IAM task role, so no static credentials touch the config. Model identifiers match the Anthropic API directly, no Bedrock ARNs or inference profiles required.

Anthropic's documentation lists Amazon Bedrock, Claude Platform on AWS, Google Cloud's Agent Platform, Microsoft Foundry, and the Anthropic API as supported upstreams. Failover between them is built in. The gateway translates the Anthropic Messages API for each backend.

Deployment choice determines the data path. With Amazon Bedrock, inference requests stay within AWS's security boundary and inherit the same data handling controls as other Bedrock workloads. Claude Platform on AWS targets teams that want Anthropic's native experience without leaving AWS authentication and billing.

The identity gap practitioners spotted immediately

Within a day of launch, engineers started probing edge cases. On LinkedIn, Christoph Klingspor asked how to handle workload identity for organizations outside the gateway's target setup: "Any thought on how we could give Claude Code an identity with a role? Best idea I had so far was AWS Private CA. Use case is for companies not yet using Anthropic on AWS but rather a normal subscription."

Shweta S., a security lead at AWS, responded with the pattern that closes the gap: "Good instinct on Private CA — but the piece that gives it an identity+role is IAM Roles Anywhere. Private CA (or your own existing CA) issues the X.509 cert; Roles Anywhere swaps it for temporary AWS creds tied to an IAM role, so no long-lived keys. Just wire up its credential helper and Claude Code's AWS tooling picks up the scoped creds automatically."

Why enterprise AI rollouts stall at cost attribution

Christopher Dorsey, an enterprise sales leader, described why the cost attribution features matter commercially: "Every enterprise AI rollout I've been around stalls in the same spot: the champion loves it, then IT and finance can't see who's spending."

That's the real problem the gateway solves. An engineering manager can approve Claude Code for their team. But without centralized controls, finance sees one giant AWS bill with no breakdown. Security sees developers with individual API keys that don't revoke when someone leaves. IT sees settings scattered across laptops. The gateway collapses all of that into one control point that plugs into existing identity providers and cost allocation systems.

The multi-cloud angle

AWS, Google Cloud, and Anthropic all publishing deployment guidance for the same control plane in the same week is unusual. Typically cloud vendors want lock-in. Here, the gateway abstracts the upstream, meaning an organization could fail over between AWS and Google Cloud backends without changing client configuration.

For teams already managing multi-cloud infrastructure with tools like Cloudflare for edge networking or running workloads across DigitalOcean and AWS, this pattern feels familiar. The control plane stays consistent; the compute backend becomes a deployment decision.

Frequently Asked Questions

Does Claude Apps Gateway require Amazon Bedrock?

No. The gateway supports multiple upstreams including Amazon Bedrock, Claude Platform on AWS, Google Cloud's Agent Platform, Microsoft Foundry, and the direct Anthropic API. You can configure failover between them.

How does the gateway handle authentication?

It acts as an OpenID Connect relying party. Developers sign in through browser-based SSO with any standards-compliant identity provider, receive short-lived tokens, and sessions automatically expire when users are removed from the identity provider.

Can I set spending limits per developer?

Yes. The gateway supports spend caps at organization, group, or individual user levels with daily, weekly, and monthly limits. Requests are blocked once a cap is exceeded.

What infrastructure do I need to run the gateway?

A single stateless container on Amazon ECS, EKS, or EC2 behind an internal Application Load Balancer. Amazon RDS for PostgreSQL stores sign-in state and rate-limit counters.

Does the gateway work with existing Claude Code installations?

Yes. The gateway ships inside the same Claude Code CLI binary developers already install. Once configured, the client picks up managed settings automatically while the gateway checks policy on every request.

Source: InfoQ