Skip to main content
Tools can opt into MCP task execution (SEP-1686) for work that outlives a normal request: the server returns a task handle immediately, runs the handler in the background, and the client polls until the result is ready. 
server.tool({ name: 'slow', description: 'A long job', task: true }, async () => {
  await doExpensiveWork()
  return 'done'
})

// fine-grained control:
server.tool({ name: 'slow', task: { mode: 'required', pollInterval: 2000 } }, handler)
This page is under construction. The outline below sketches what it will cover.
  • Task modes — forbidden (default: task requests rejected), optional (task: true: both sync and task execution work), required (sync calls rejected; advertised with execution: { taskSupport: 'required' }).
  • Execution flow — server returns CreateTaskResult immediately; client polls tasks/get until terminal, then fetches output via tasks/result. ToolConfig.timeout is intentionally not applied in task mode.
  • Stores — InMemoryTaskStore is used automatically; supply a custom TaskStore via new FastMCP({ tasks: { store: myStore } }), or enable the capability globally with { tasks: true }.
  • Error model — server-side rejections and failed handlers surface as { type: 'error' } messages in task streams rather than thrown exceptions.