Claude Code gets a self-hosted control plane as AWS ships a deployment for the apps gateway
Anthropic's Claude apps gateway ships inside the Claude Code CLI and puts SSO, spend caps, model policy, and telemetry in front of Claude Code and Claude Desktop. AWS added a reference deployment on ECS, EKS, or EC2.
Anthropic has added a self-hosted control plane to Claude Code and Claude Desktop, and AWS has now published a reference deployment for running it on its own infrastructure. The gateway ships inside the Claude Code CLI binary and sits between developers and the model, handling single sign-on, spend caps, model and tool policy, and usage telemetry. It is a quiet but telling shift: the interesting part of a coding agent for a large organization is no longer only the model, it is the layer that decides who can run it, on what, and for how much.
What shipped
The Claude apps gateway is a stateless service an organization runs itself. Developers authenticate through a standards-compliant OpenID Connect provider, and the gateway issues a short-lived token the CLI uses for every subsequent request. Sessions refresh silently, and removing someone from the identity provider expires their access within the configured lifetime, one hour by default. That removes a persistent enterprise headache, which is developers carrying long-lived personal API keys that never quite get rotated or revoked.
The governance surface is where the product earns its name. Administrators set spend limits by day, week, or month at the organization, group, or user level, and once a developer hits their cap the gateway blocks further requests until the period resets or an admin raises it. Policy controls extend to which models a group can call, which tools are permitted such as file writes or web access, and which environment defaults get distributed. Requests route to Amazon Bedrock or the Claude Platform on AWS, with cross-region and cross-account failover, and the model IDs match the public Anthropic API, so claude-sonnet-5 and claude-opus-4-8 behave the same as they would elsewhere. Every request is stamped with a usage metric that the gateway relays over OpenTelemetry to a collector the organization chooses, such as CloudWatch or Amazon Managed Service for Prometheus.
AWS’s contribution is the deployment story. Its guide runs the gateway as a single container on Amazon ECS, EKS, or EC2, backed by Amazon RDS for PostgreSQL for session state and rate-limit counters and fronted by a load balancer with managed TLS. Because the gateway holds the upstream credentials, individual developers never touch cloud keys directly.
Where this lands in the market
The framing worth holding onto is that coding agents are moving from individual tools to governed infrastructure, and the vendors appear to be racing to own that governance layer rather than leave it to third parties. The past few weeks made the stakes concrete. Alibaba banned Claude Code internally after finding undisclosed fingerprinting behavior, a US state moved to standardize on Claude across its agencies, and a Canadian province ran hundreds of millions of lines of government code through Claude Code agents. Each of those decisions hinges less on model quality, which is converging across vendors, and more on whether an organization can prove who ran what and cap the bill before it surprises finance.
For a working developer inside one of these companies, the practical change is that authentication moves to the corporate SSO login they already use, spending is bounded by policy rather than trust, and the choice of model and tools may no longer be theirs to make. That is a loss of some individual autonomy in exchange for a version of the agent that a security and finance team will actually approve. The pattern also generalizes: a self-hosted control plane that fronts a coding agent is exactly the shape GitHub, and by extension its competitors, will need if agents are going to run at company scale rather than on individual laptops.
What’s worth watching
The open question is whether a self-hosted gateway becomes the default way enterprises run coding agents, or stays a checkbox for the most regulated buyers while everyone else keeps using managed cloud endpoints. The signal to watch is adoption friction: a stateless container plus a managed database is a modest lift for an AWS-native team, but the gateway still has to be maintained, monitored, and kept current with the CLI it ships inside.
The broader thread is that model access is becoming a policy problem as much as a capability one. Over the next 90 days, the things to track are whether other clouds publish equivalent deployments, whether Anthropic extends the same controls to its consumer and web surfaces, and how spend-cap enforcement holds up once real teams push agent workloads through it. Stackmaven’s follow-up coverage will land on or around October 16.