Table of Contents

Tasks

Tasks let an MCP server run a request asynchronously and report its result to the client later. The primary use case today is long-running tool invocations: the tool is offloaded to a background task, and the client polls for status, optionally exchanging additional input along the way.

Tasks are provided by the ModelContextProtocol.Extensions.Tasks package and require MCP protocol version 2026-07-28 or later. The implementation follows SEP-2663 (Tasks Extension).

Overview

A client opts into tasks on a per-request basis by including the io.modelcontextprotocol/tasks extension key in the request's _meta. When that opt-in is present, the server might respond with a CreateTaskResult instead of the standard result (for example, CallToolResult). The client then polls tasks/get until the task reaches a terminal state.

Per the SEP, the server must not return CreateTaskResult for a request that did not include the extension opt-in. The SDK enforces this on the server side.

Task lifecycle

                  ┌─────────────────────────┐
                  ▼                         │
   (start) → Working ──→ InputRequired      │
              │              │              │
              │              └──────────────┘  (client responds via tasks/update)
              │
              ├──→ Completed   (terminal — includes tool results with isError: true)
              ├──→ Cancelled   (terminal)
              └──→ Failed      (terminal — JSON-RPC errors only)

McpTaskStatus wire values are serialized in snake_case: working, input_required, completed, cancelled, failed.

The discriminator field Result.ResultType on the response payload is "task" for CreateTaskResult and "complete" for ordinary results.

Server configuration

Using the task store

The easiest way to enable the package's task store integration is to call WithTasks on the server builder, passing an IMcpTaskStore. The SDK ships InMemoryMcpTaskStore for development and tests:

using ModelContextProtocol.Extensions.Tasks;

builder.Services.AddMcpServer()
    .WithTools<MyTools>()
    .WithTasks(new InMemoryMcpTaskStore());

When tasks are enabled with WithTasks the SDK automatically:

  • Wires the tasks/get, tasks/update, and tasks/cancel handlers from the store.
  • Advertises the io.modelcontextprotocol/tasks extension in ServerCapabilities.Extensions.
  • Wraps each [McpServerTool] invocation so that, when the client opts in to the extension, the tool is offloaded to a background task tracked by the store.
  • Establishes a task scope so that ElicitAsync, SampleAsync, and RequestRootsAsync called from inside the tool surface as entries in the task's inputRequests instead of as direct JSON-RPC requests.
  • Plumbs a CancellationToken through to the tool that fires when the client invokes tasks/cancel, so cancellation propagates cooperatively.

Alternate-result tools/call filters run in registration order, with the Tasks filter creating a task at its position in that order. Filters before Tasks run before task creation. Filters after Tasks run in the background before the ordinary filter pipeline. ASP.NET Core tool authorization uses an alternate-result filter registered before Tasks, so an unauthorized call does not create a task.

Ordinary tools/call filters still run exactly once for task-backed calls. They execute in the background after the task record is created and before the tool body, so validation and telemetry continue to apply. Each background invocation gets an independent DI scope that remains alive until the tool pipeline completes.

For production scenarios that need durability, session isolation, multi-process routing, or TTL-based cleanup, implement IMcpTaskStore yourself (see the Implementing a custom task store section).

Returning a task from a tool handler

For full control without the store's auto-wrapping, set McpServerHandlers.CallToolWithAlternateHandler. It returns a ResultOrAlternate<TResult>, so each invocation can choose between an immediate result and an alternate result such as a CreateTaskResult:

using ModelContextProtocol.Extensions.Tasks;

options.Handlers.CallToolWithAlternateHandler = async (context, ct) =>
{
    if (ShouldRunInline(context.Params!))
    {
        return new CallToolResult { Content = [/* … */] };
    }

    var taskId = await StartBackgroundWorkAsync(context.Params!, ct);
    var created = new CreateTaskResult
    {
        TaskId = taskId,
        Status = McpTaskStatus.Working,
        CreatedAt = DateTimeOffset.UtcNow,
        LastUpdatedAt = DateTimeOffset.UtcNow,
        PollIntervalMs = 1000,
    };

    return new ResultOrAlternate<CallToolResult>(created, McpTasksJsonContext.Default.CreateTaskResult);
};

This low-level handler is mutually exclusive with WithTasks. When a store is configured, the SDK does the wrapping for you and throws InvalidOperationException if the alternate handler also returns an alternate. Use one mechanism or the other. When you return a task this way, you're also responsible for serving tasks/get, tasks/update, and tasks/cancel, which the store provides automatically.

McpServerHandlers.CallToolHandler and McpServerHandlers.CallToolWithAlternateHandler are mutually exclusive. Setting one while the other is already non-null throws InvalidOperationException at the property setter.

Client usage

Automatic polling

CallToolWithPollingAsync handles the full task lifecycle automatically:

  • Injects the io.modelcontextprotocol/tasks extension capability into the request's _meta.
  • Polls tasks/get at the cadence the server suggests via pollIntervalMs.
  • Dispatches input requests through the client's registered handlers (SamplingHandler and ElicitationHandler).
  • Deduplicates already-resolved input request keys across polls so each request is handled at most once.
  • Returns the final CallToolResult when the task completes, or throws McpException on Failed/Cancelled.
using ModelContextProtocol.Extensions.Tasks;

var result = await client.CallToolWithPollingAsync(
    new CallToolRequestParams { Name = "long-running-tool", Arguments = arguments },
    cancellationToken: cancellationToken);

Manual control

Use CallToolAsTaskAsync to receive the raw ResultOrCreatedTask<TResult> without auto-polling, then drive the lifecycle yourself using GetTaskAsync, UpdateTaskAsync, and CancelTaskAsync:

using ModelContextProtocol.Extensions.Tasks;

var raw = await client.CallToolAsTaskAsync(requestParams, cancellationToken);
if (raw.IsTask)
{
    var taskId = raw.TaskCreated!.TaskId;
    while (true)
    {
        await Task.Delay(TimeSpan.FromMilliseconds(raw.TaskCreated.PollIntervalMs ?? 1000), cancellationToken);
        var state = await client.GetTaskAsync(taskId, cancellationToken);
        // Handle InputRequiredTaskResult by calling UpdateTaskAsync,
        // CompletedTaskResult by deserializing its Result property, etc.
    }
}

Stuck-task detector

CallToolWithPollingAsync includes a safety net for misbehaving servers: if the task stays in InputRequired across many consecutive polls without exposing any new input request keys (that is, every previously requested input has already been resolved by the client and yet the server keeps returning InputRequired), the client gives up, issues a best-effort tasks/cancel, and throws McpException. This guards against a server that never transitions out of InputRequired and prevents an unbounded poll loop.

The threshold defaults to 60 consecutive stuck polls and is configurable via the maxConsecutiveStuckPolls parameter on CallToolWithPollingAsync. The effective wall-clock timeout is roughly maxConsecutiveStuckPolls * pollIntervalMs, so tune the value with the server-side poll cadence in mind. Setting it too low risks false positives for servers that are slow to surface follow-up input requests; setting it too high can mask misbehaving servers.

Input requests (multi-round-trip)

When a task needs additional input from the client, the server transitions it to InputRequired and returns the outstanding requests in InputRequests. Each entry is an arbitrary key paired with a { method, params } envelope representing an equivalent standalone server-to-client request. The client provides answers via UpdateTaskAsync, keyed by the same identifiers.

Supported input request methods:

Method Dispatched to the client handler
elicitation/create ElicitationHandler
sampling/createMessage SamplingHandler

Per SEP-2663:

  • Each input request key must be unique over the lifetime of the task.
  • Clients should deduplicate keys across polls so a request is only presented to the user or model once. CallToolWithPollingAsync does this automatically.
  • Servers should ignore inputResponses entries whose key does not currently correspond to an outstanding request, including responses for terminal-state tasks. InMemoryMcpTaskStore follows this rule.

Implementing a custom task store

Implement IMcpTaskStore for production scenarios. Key requirements drawn from the SEP and the SDK contract:

  1. Thread safety — every method can be called concurrently.
  2. Idempotent terminal transitions — SetCompletedAsync, SetFailedAsync, and SetCancelledAsync must be no-ops on a task that is already in a terminal state so a late cancellation cannot overwrite a result.
  3. InputResponseReceived event — after persisting an input response inside ResolveInputRequestsAsync, raise IMcpTaskStore.InputResponseReceived for each resolved entry. This is the only mechanism that wakes a pending server.ElicitAsync/server.SampleAsync call waiting inside a task scope. In distributed deployments where a different server instance receives the tasks/update, the event must be propagated to the originating server (for example via Redis pub/sub, SignalR, or a custom transport).
  4. Strong-consistency on CreateTaskAsync — CreateTaskAsync must not return until the task is durably persisted, so that a subsequent GetTaskAsync with the returned task ID resolves immediately — even from a different process or node. Stores backed by eventually consistent storage must wait for the write to become visible (quorum acknowledgement, write-through, etc.) before returning. Required by SEP-2663 §306.
  5. Singleton under stateless HTTP — when the server runs in stateless mode (each request spins up a fresh server instance), the same IMcpTaskStore instance must be shared across requests — either by registering it as a singleton in DI, or by backing it with external storage that every instance can reach. Otherwise tasks/get polls from subsequent requests will see an empty in-memory store and never find the task.
public sealed class MyTaskStore : IMcpTaskStore
{
    public event Action<InputResponseReceivedEventArgs>? InputResponseReceived;

    public async Task ResolveInputRequestsAsync(
        string taskId,
        IDictionary<string, InputResponse> inputResponses,
        CancellationToken cancellationToken = default)
    {
        // 1. Atomically persist the resolved requests, ignoring keys that are no longer
        //    outstanding or that target a terminal task.
        await PersistResolvedResponsesAsync(taskId, inputResponses, cancellationToken);

        // 2. Then notify subscribers so any awaiting server.ElicitAsync/SampleAsync resumes.
        foreach (var kvp in inputResponses)
        {
            InputResponseReceived?.Invoke(new InputResponseReceivedEventArgs
            {
                TaskId = taskId,
                RequestId = kvp.Key,
                Response = kvp.Value,
            });
        }
    }

    // … other IMcpTaskStore members
}

Status semantics

Completed is the terminal status whenever the underlying request produced its standard result, including a CallToolResult with IsError = true. Per SEP-2663, tool-level error results are not promoted to Failed.

Failed is reserved for JSON-RPC protocol-level errors during execution — for example, a malformed request, or an unhandled exception in a custom handler that the SDK converts to a JSON-RPC error. Use CallToolResult.IsError for domain-level errors the model should see.

Cancellation semantics

Per SEP-2663, tasks/cancel is eventually consistent and cooperative: the server acknowledges the request immediately, but is not required to actually stop the work or to transition to Cancelled. The notifications-cancelled mechanism (used for plain JSON-RPC requests) is not used for task cancellation; clients must use tasks/cancel.

In the built-in SDK pipeline, when a task is wrapped by a configured TaskStore:

  1. The store's SetCancelledAsync transitions the task to Cancelled (a no-op if the task is already terminal).
  2. The associated CancellationTokenSource is signaled, propagating cancellation to the tool's CancellationToken so cooperative cleanup can run.
  3. Whichever side (the cancel handler or the background runner's finally block) wins TryRemove on the cancellation source owns disposal, avoiding ObjectDisposedException.

Architecture notes

Immutable store design

InMemoryMcpTaskStore uses immutable record snapshots with compare-and-swap updates for lock-free thread safety. InputRequests and InputResponses are exposed as ImmutableDictionary<,> so observers can't mutate internal state.

Capability bypass inside a task scope

When server.ElicitAsync/server.SampleAsync/server.RequestRootsAsync execute inside a task scope, the SDK intentionally skips the normal client-capability negotiation checks (ThrowIfElicitationUnsupported, etc.). The tasks extension itself is the negotiated capability: the client opted in by including the extension marker in the originating request, so it's responsible for handling — or rejecting — the input requests surfaced through tasks/get.

Compatibility with v1 experimental Tasks

The Tasks extension in v2.0.0 replaces the experimental Tasks implementation shipped in v1.3.0 and v1.4.x. The implementations are not compatible at either the API or protocol level. A v2 Tasks client or server must use a connection negotiated to 2026-07-28 or later; it cannot fall back to the down-level implementation.

On a connection negotiated to the down-level 2025-11-25 protocol:

  • A v2 client calling a v1 server receives an ordinary tool result. The v2 client does not opt in to the down-level Tasks protocol, and GetTaskAsync rejects use before a 2026-07-28 connection is negotiated.
  • A v1 client calling a v2 server likewise receives an ordinary tool result. The v2 server does not create Tasks on a down-level connection, and its tasks/get endpoint rejects the legacy request with a method-not-found error.

Upgrade both peers to the v2 Tasks extension before using Tasks. The extension provides no compatibility bridge for the previous experimental API.

Known limitations

  • Server-push task status notifications (SEP-2575): not yet implemented. Clients rely on polling exclusively.
  • Lazy task creation: when a tool runs through the task store, the store's CreateTaskAsync is invoked eagerly before the inner handler runs, so tools that complete inline still incur a store write. There is currently no built-in deferral.
  • Mid-execution promotion to task: an [McpServerTool] method cannot start executing synchronously and then transition its remaining work to a background task. Use a custom McpServerHandlers.CallToolWithAlternateHandler if you need that pattern.
  • roots/list as an input request: the server SDK routes RequestRootsAsync through the task channel when called from inside a task scope, but the client SDK does not currently dispatch a handler for that method. Avoid calling server.RequestRootsAsync from within a task scope until client-side support is added.
  • ServerCapabilities.Extensions round-trip: the dictionary is typed as IDictionary<string, object> so its values cannot be deserialized by the source generator. The negotiated extension surfaces correctly at the wire level, but round-tripping arbitrary extension payloads in-process is not supported.