On the Microsoft Foundry blog, Naveen Gopalakrishna published an engineering guide to Claude Code on Microsoft Foundry with Azure API Management, built around the question every platform team is asking: how do you give developers Claude Code “while keeping identity, model access, budgets and private networking under enterprise control?” The answer, which he calls CCOF, puts API Management in front of the Claude deployments. Developers sign in with Entra and get a short-lived gateway token instead of a shared Foundry key; the gateway validates tenant, audience, client and scope, asks a resolver function which group and models the caller is entitled to, applies per-person and per-group token limits, and forwards to a private Foundry backend.
Two labels matter before anyone copies it. CCOF “is my implementation” in the author’s GitHub repository, “not a Microsoft product,” even though it runs on a Microsoft blog. And the numbers are lab numbers, stated as such: a 2–3 hour deployment window, about 2 hours 12 minutes to the priced API goal and 2 hours 28 minutes to the first real Claude Code reply, and a configurable 120-second allow cache. What survives is the design rule, “Access comes from group membership. A client configuration file does not grant access by itself,” and the hard-won networking advice: private IPs belong in DNS, not in the backend URL, or TLS validation breaks.
The gateway in this architecture is a catalog provider. The Microsoft Azure API Management provider page lists 133 API pages, and its record already describes an AI gateway governing Anthropic Messages and Microsoft Foundry models with token controls. The guide’s moving parts map onto specific pages: the approved Foundry endpoints are registered through the Backend API, admission and token limits are enforced through the Policy API, and the person, group and model attribution behind chargeback rides on the Logger API.
API Management scores 80.3 on the Kin Score, exemplar, with access clarity at 92.1, developer ergonomics at 91.1 and operational transparency at 86.8. Agent Readiness is 31.8, agent-ready, with auth clarity, error semantics and idempotency lit. Delegated identity is unlit, as are consent identity and protected resource metadata. That is the gap the guide fills by hand: the entire design is a developer’s identity delegated to an agent through a gateway, and it took a resolver, a sync function and a reconciler to say who may act. The gateway can enforce delegation; its own contract does not yet describe it.