Stores

Every durable store implements the same contract as the in-memory store in Authenticated tasks: stamp the owner on create, isolate client-facing get/cancel/resolve to that user, let the SDK complete work after the HTTP request is gone. Pick the package that matches the engine already in your host. The identity factory stays the same when you swap. The design lives in specs/authenticated-task-stores.md.

In-process

Package	How-to	When
`McpEssentials.Core`	Authenticated tasks	Tests, local MCP, a single process

InMemoryMcpTaskStorePlus keeps snapshots in a ConcurrentDictionary. Fan-out is an in-process event. Tasks die with the process.

Database

Relational and document engines. Entity Framework Core is the default durable package: one implementation, any EF Core provider.

Package	How-to	Engine	Status
`McpEssentials.EntityFrameworkCore`	Entity Framework Core	SQL Server, SQLite, PostgreSQL, MySQL, Cosmos via EF, and any other EF Core provider	Shipped
`McpEssentials.PostgreSQL`	PostgreSQL	Npgsql, no EF	Specified
`McpEssentials.SqlServer`	SQL Server	`Microsoft.Data.SqlClient`, no EF	Specified
`McpEssentials.Cosmos`	Cosmos DB	Azure Cosmos DB for NoSQL	Specified
`McpEssentials.MongoDB`	MongoDB	MongoDB	Specified
`McpEssentials.AzureTables`	Azure Tables	Azure Table Storage	Specified

Use EF Core when the host already has a DbContext. Native PostgreSQL and SQL Server packages exist for hosts that want ADO.NET (and, for Postgres, LISTEN / NOTIFY) without an EF model.

Distributed cache

Low-latency handles with native or host-owned expiry.

Package	How-to	Engine	Status
`McpEssentials.Redis`	Redis	StackExchange.Redis, `EXPIRE` plus pub/sub	Specified
`McpEssentials.DistributedCache`	IDistributedCache	Whatever you already registered as `IDistributedCache`	Specified

Prefer Redis when you need pub/sub for elicitation across nodes. Use IDistributedCache when that abstraction is already in DI and a host-supplied bus covers fan-out.

Eventing

Stores whose native bus is the reason the package exists.

Package	How-to	Engine	Status
`McpEssentials.Nats`	NATS	JetStream KV for the row, Core NATS for signals	Specified

Providers without a native bus (EF Core, IDistributedCache, Azure Tables) take an optional IMcpTaskInputSignalBus. Missing bus plus multiple nodes means ElicitAsync waits until timeout.

Shared row

StoredMcpTask is the persistence foundation. Durable entities subclass it (EF adds RowVersion). Keep these rows on their own table; tasks expire in minutes.

Field	Role
`TaskId`	Primary key
`OwnerUserId`	Isolation key
`Status`	`Working` / `InputRequired` / `Completed` / `Cancelled` / `Failed`
`CreatedAt` / `LastUpdatedAt`	UTC
`TimeToLive` / `ExpiresAt`	Work window (`CreatedAt + TimeToLive`)
`DiscardAfter`	Visibility ceiling (`CreatedAt + 2×TimeToLive`), stamped at write
`ResultJson` / `ErrorJson` / `InputRequestsJson`	Payloads

Swap in production

HTTP identity stays on the platform accessor. Development uses Core. Production registers a durable store with the same factory.

builder.Services.AddHttpContextAccessor();

Func<IServiceProvider, ClaimsPrincipal?> currentUser =
    sp => sp.GetRequiredService<IHttpContextAccessor>().HttpContext?.User;

// Development
builder.Services.AddMcpServer()
    .WithHttpTransport(o => o.Stateless = true)
    .WithTasksPlus(currentUser, o => o.DefaultTimeToLive = TimeSpan.FromMinutes(5));

// Production (any EF Core provider on AppDbContext)
builder.Services.AddDbContext<AppDbContext>(o => o.UseSqlServer(connectionString));
builder.Services.AddMcpServer()
    .WithHttpTransport(o => o.Stateless = true)
    .WithTasksPlus<AppDbContext>(currentUser, o => o.DefaultTimeToLive = TimeSpan.FromMinutes(5));

Map McpTasks with AddMcpTasksTable. See Entity Framework Core for which overload to call (OnModelCreating vs AddDbContext). Isolation helpers stay in Core. Provider packages depend on McpEssentials.Core, not on each other.

Getting Started

API Reference

In-process

Database

Distributed cache

Eventing

Shared row

Swap in production

​In-process

​Database

​Distributed cache

​Eventing

​Shared row

​Swap in production

In-process

Database

Distributed cache

Eventing

Shared row

Swap in production