Chapter 1 of 6
Module scope#
What we will do#
Put an identity-controlled gateway between Claude developer tools and Foundry. Configure one native Anthropic API in Azure API Management, then connect Claude Code and Claude Desktop. An assigned user receives a streaming model response; APIM rejects an invalid identity or an unassigned model before calling Foundry.
Why it matters#
Developer-managed provider keys make individual access and spend difficult to control. APIM validates the signed-in user, applies a monthly model budget, and uses its own managed identity to call Foundry.
Boundaries#
This optional module extends Control live AI traffic, beside the numbered sessions. It configures an existing nonproduction APIM instance, two Entra applications, and managed client settings. Live Azure and Entra state remains authoritative. This route invokes Claude models directly; it does not deploy a Foundry Agent Service agent or govern local tools and MCP calls. The operations owner takes over usage monitoring and client rollout.
Module preparation
Who should join
- API Management and Microsoft Foundry platform engineers
- Microsoft Entra identity and endpoint-management engineers
- Security, finance, and developer-tool owners
What you need
- An approved nonproduction API Management v2 instance has a system-assigned managed identity.
- The Foundry owner has deployed approved Claude models, accepted the Marketplace terms, and confirmed the deployment names, hosting location, and region scope.
- The gateway operator has time-bound Contributor on the exact API Management resource group and Reader on the Foundry resource for preflight.
- The role-assignment owner has User Access Administrator on the Foundry resource and can grant Cognitive Services User to the APIM identity.
- An Application Administrator can configure two single-tenant Entra applications, consent, app roles, and assigned users.
- An endpoint-management owner can distribute Claude Code settings and an exported Claude Desktop third-party inference profile.
- The gateway owner has approved the public or private network path, body-free Application Insights diagnostics, and the restore change.