The real test for any AI coding tool is not whether it impresses a developer on a personal project. It is whether a mid-size enterprise can route it through their existing identity systems, cap access by team, log every token for compliance, and swap the underlying model without rewriting their toolchain. That is the actual bar. And it is the bar that Anthropic’s Claude Apps Gateway is now trying to clear.
The gateway is a self-hosted proxy that sits between your organization and Claude Code. It runs on your own infrastructure. Developers authenticate through SSO rather than through individual Anthropic accounts. Groups get different levels of model access. Telemetry flows out through OTLP, which is the standard observability pipeline most engineering organizations already have. This is not a product announcement that reads like a demo. It reads like something an IT security team actually asked for.
What makes this worth watching is the deployment surface it covers. The gateway can route Claude through Amazon Bedrock, Google Cloud, or Microsoft Foundry, meaning enterprises do not have to touch Anthropic’s API directly if they do not want to. That matters to any company with a cloud commitment or a data residency requirement. The model becomes, in a real sense, interchangeable; the gateway layer stays the same. That is a familiar pattern from how enterprise software works, and it is not by accident that Anthropic is building toward it.
Microsoft is on the other end of this. Azure Foundry now hosts Claude models including Mythos, Fable, Opus, Sonnet, and Haiku, which means the full model family is available inside an Azure tenant. For a company already running its developer tooling inside Azure, adding Claude Code through Foundry is the path of least resistance. The procurement is already done. The compliance is already handled. The only question is whether the coding tool itself is good enough to justify the rollout.
This is the same problem every enterprise AI deployment runs into. The model quality question gets answered relatively quickly in a proof of concept. The deployment question takes six months and involves legal, security, procurement, and whoever controls the SSO configuration. Anthropic is building infrastructure for that second problem, not the first one. That is a different strategic posture than where the company was even a year ago.
The per-group model access control is the detail I keep coming back to. It means an organization can give its senior engineers access to Opus while routing everyone else to Sonnet or Haiku based on cost or capability requirements. That is not a feature a developer needs. That is a feature a VP of Engineering needs when trying to justify the budget line. It is also the kind of control that makes the difference between a pilot that stays a pilot and a deployment that actually scales.
The OTLP telemetry support is doing similar work. OTLP is the OpenTelemetry Protocol; it is how modern observability stacks collect traces and metrics. If Claude Code emits telemetry in a format that plugs into Datadog, Grafana, or whatever the organization already uses, then the AI coding tool becomes observable in the same way any other service is observable. You can see where it is slow, where it is expensive, where it is being used and by whom. That turns a black box into something an engineering organization can actually manage.
Anthropic has been public about wanting to build for enterprises without sacrificing the developer experience that made Claude Code appealing in the first place. The gateway approach tries to thread that needle by keeping the developer-facing surface the same while adding the enterprise control plane underneath it. Whether that works in practice depends on how smooth the SSO integration actually is and how much configuration the self-hosted setup demands before it is usable. Documentation says the setup covers sign-in, group routing, and telemetry; the real question is how many hours of an infrastructure engineer’s time it takes to get there.
The Microsoft Foundry integration is also an interesting signal about where the distribution fight for enterprise AI is heading. OpenAI has Azure as its primary enterprise distribution channel, which gives it enormous reach inside companies already committed to Microsoft. Anthropic getting Claude into Foundry as a first-class option, with the full model family available, is a direct play for the same distribution surface. The question is not whether enterprises will use AI coding tools; they will. The question is which model ends up in their default environment, and right now the answer to that question is shaped less by benchmark performance than by which API is already approved, already contracted, and already inside the firewall.
The documentation for the gateway describes it as enabling “per-group model access, and OTLP telemetry,” which is about as dry a product description as you can write and also exactly what enterprises need to hear. It is the opposite of how AI products usually get announced. There is no mention of how the model thinks or feels about code. There is just a description of the control surface and how it connects to the infrastructure you already have.
This blog has tracked how the real adoption story in AI is almost always downstream of the deployment story; the model that wins is not always the best model, it is the model that gets past the IT review.