Appearance
Model Router, gateway and MCP roles
The deployed pattern has two Foundry accounts, each containing its own project, and one gateway. Projects are children of accounts, not containers for multiple accounts. Regional placement follows model availability and quota. Exact names and production verification belong in the private overlay.
Two kinds of routing
MCP roles map jobs to deployments. The calling assistant chooses a role; its description guides the caller and is not an Azure task classifier. The auto role maps to Azure Model Router. The gateway forwards that request to the hosting account, where Azure chooses an eligible underlying model. Other roles and direct model selection bypass automatic selection.
| Request | Model selection |
|---|---|
MCP code, docs, iac or another role | Private MCP routing policy |
| MCP direct deployment | Caller, subject to MCP allow/block policy |
MCP auto or direct router deployment | Azure selects within its allowed pool |
| Editor model picker | User; Azure selects further only for the router entry |
The current deployment uses router version 2025-11-18, GlobalStandard, Balanced mode, and an explicit GPT-5.6 Luna, Terra and Sol subset. fast remains the MCP default. Capacity is a SKU-specific quota allocation, not a context window or a universal token-per-minute value.
Pool configuration
Microsoft supports Balanced, Cost and Quality modes and custom subsets. Use only models supported by the router version and region. Deploying GPT-6, Kimi or DeepSeek V4 separately does not make them eligible for the current pool. Read back effective routing settings after updates, rather than relying only on provisioning success. Microsoft configuration reference.
Use explicit roles when a task must use a particular model. Evaluate Cost mode against Balanced on fixed representative prompts before adopting it. Cost mode is not a spending cap; monitor selected models and billable tokens, with budgets and output limits as separate controls. Microsoft evaluation guidance.
Using MCP
The local adapter offers list_models, ask_model and compare_models; the hosted platform exposes corresponding Foundry tools. List models first. Supply model: "auto" to the ask tool for Azure routing. Omitting the model uses the configured default role. Include required context in the prompt: delegated models do not inherit files, tools or the caller's conversation.
The private overlay owns its registry and MCP routing.json. Sync-McpRouting.ps1 generates hosted platform configuration and supports -Check for drift detection. A hosted policy change requires a platform commit, merge, build/release and live verification. Local JSON edits alone do not update the remote server. Block rules also apply to direct selection.
Monitoring and recommendations
Inspect requested versus returned model, backend, errors, latency, token coverage and cost freshness. First-token timing requires streaming; normal non-streaming MCP requests cannot populate it. A successful router response proves connectivity, not task quality or cost optimality.
The dated research led to the owner-approved September 25 role update: GPT-6 Luna for fast/docs/cheap-bulk, Kimi K2.7 Code for code, Grok 4.6 for deep/adversary, GPT-6 Sol for iac, DeepSeek V4 Pro for second-opinion, and the existing Balanced router for auto. Escalation requires an explicit caller decision; there is no automatic fallback chain. The private overlay records the hosted release and live verification. See the cloud roster, gateway guide and monitoring operations.