Custom Model Endpoint, Model Selection, and MCP Connector Reuse in Microsoft 365 Copilot

We're evaluating an integration between Microsoft 365 Copilot and a Claude model we host on Microsoft Foundry, and need to confirm current product capability before finalizing our architecture.

1. Direct Foundry Endpoint in Native M365 Copilot

Is there any supported way - today, in preview, or on a near-term roadmap to configure Microsoft 365 Copilot's native experience (Business Chat, and the in-ribbon Copilot entry point in Word, Excel, Outlook, and PowerPoint) to route to a model hosted in our own Microsoft Foundry resource, rather than Microsoft's default hosted models?

We're aware custom engine agents published via Teams/Copilot Chat exist as a separate, user-invoked surface, we need to confirm specifically whether the native Copilot experience itself can be pointed at a customer-managed endpoint, at any tenant or admin level.

Additionally, can we use the custom engine agents in Office Apps (If we have M365 Copilot License)?

2. Model Picker in the M365 Copilot Interface

Is there any existing or planned capability for end users, or tenant admins via policy, to select which underlying model handles a given Copilot session directly within the Copilot chat interface? If anything along these lines exists in preview or is on a roadmap, we'd appreciate details.

3. Reuse of Tenant MCP Connectors by Pro-Code Custom Engine Agents

There are MCP servers already registered as Microsoft 365 Copilot connectors at the tenant level, but pro-code custom engine agents (built via the Microsoft 365 Agents SDK or Microsoft Foundry) do not inherit that same connector registry and must instead independently configure access to the same MCP Server connections.

Can you confirm whether this is accurate, and whether there is any supported way for a Foundry-built custom engine agent to reuse an MCP server already registered as a tenant-level Copilot connector, rather than reconfiguring that same connection separately?

We'd like to confirm whether any of the above is achievable through configuration, licensing, or preview features before deciding the feasibility.