Long-running actions - EasyAF.AI

Most calls finish when your API returns. Shipping an order, kicking off a report, or starting a sign-in that opens a browser can take longer than one chat turn. Put [McpTask] on that action. The harness gets a handle, polls until you are done, and then sees the real result. CRUD, queries, and unmarked actions still finish immediately.

Mark the action

Restier / controllers
ModelBuilder

using Microsoft.OData.Mcp.Core.Vocabularies;

[BoundOperation(OperationType = OperationType.Action)]
[McpTask]
public Order Ship(Order order) => order;

builder.EntityType<Order>().Action("Ship").HasMcpTask().Returns<Order>();

$metadata will show the MCP Task term on that action. Named action_order_ship and odata_call of Ship behave the same.

Turn waiting on

builder.Services.AddODataMcp(options => { }, mcp =>
{
    mcp.WithTasks(new InMemoryMcpTaskStore());
});

That store is in-memory and single-process. If the app scales out, give WithTasks a store every instance can see. UseODataMcp() stays as it is. Waiting is configured on AddODataMcp. If the action also returns a URL the operator must open, add elicitation and a waiter so “done” means the grant finished, not merely that the link was shown.

What the harness actually sees

When it opted into Tasks, tools/call returns a task id. It polls until the work completes or is cancelled. Without that opt-in, the same annotated action is still one response with whatever JSON you returned (usually a Started ticket).