Publish multi-version documentation site (#1727)

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: bb4b44f0-f505-49c5-9878-76e2b40c7a15
Copilot-Session: 8eac1204-10ae-4319-9333-aee5dd43dcc8
This commit is contained in:
Jeff Handley 2026-07-24 15:55:35 -07:00 • committed by GitHub
commit ae36026695
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -28,11 +28,11 @@ When sessions are enabled (`Stateless = false`), the server creates and tracks a
## Forward and backward compatibility
The `Stateless` property is the single most important setting for forward-proofing your MCP server. The default is now `Stateless = true` (sessions disabled), which is the forward-compatible setting for the `2026-07-28` protocol revision and beyond. Stateless servers still respond to clients on `2025-11-25` and earlier — the SDK keeps the `initialize` + `Mcp-Session-Id` handshake available for those clients — but they cannot use the session-dependent features ([unsolicited notifications](#how-streamable-http-delivers-messages), resource subscriptions, per-client isolation). Server-to-client requests are the exception: [elicitation](xref:elicitation) — and the now-deprecated [sampling](xref:sampling) and [roots](xref:roots) — can run statelessly through [MRTR](xref:mrtr), though MRTR requires the unratified `2026-07-28` revision and is far less widely supported than session-based requests. We recommend every server set `Stateless` explicitly rather than relying on the default:
The `Stateless` property is the single most important setting for forward-proofing your MCP server. The default is now `Stateless = true` (sessions disabled), which is the forward-compatible setting for the `2026-07-28` protocol revision and beyond. Stateless servers still respond to clients on `2025-11-25` and earlier — the SDK keeps the `initialize` + `Mcp-Session-Id` handshake available for those clients — but they cannot use the session-dependent features ([unsolicited notifications](#how-streamable-http-delivers-messages), resource subscriptions, per-client isolation). Server-to-client requests are the exception: [elicitation](xref:elicitation) — and the now-deprecated [sampling](xref:sampling) and [roots](xref:roots) — can run statelessly through [MRTR](xref:mrtr) when both peers support `2026-07-28`. We recommend every server set `Stateless` explicitly rather than relying on the default:
- **`Stateless = true`** — the current default and the forward-compatible choice. Your server opts out of sessions entirely and the `Mcp-Session-Id` header is never sent or used. The `2026-07-28` protocol revision drops the `initialize` handshake and `Mcp-Session-Id` from the wire format entirely, so this is the only configuration that lets the server respond to `2026-07-28` clients without falling back to initialize-handshake handling. If you don't need [unsolicited notifications](#how-streamable-http-delivers-messages), server-to-client requests, or session-scoped state, this is the setting to use today.
- **`Stateless = false`** — the right choice when your server depends on sessions for [unsolicited notifications](#how-streamable-http-delivers-messages), resource subscriptions, or per-client isolation, none of which work without a session. Setting this explicitly protects your server from a future default change, and the [MCP specification requires](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports#streamable-http) that clients use sessions when a server's `initialize` response includes an `Mcp-Session-Id` header, so compliant clients always honor your server's session. Server-to-client requests no longer force a session: [elicitation](xref:elicitation) — and the now-deprecated [sampling](xref:sampling) and [roots](xref:roots) — can run statelessly through [MRTR](xref:mrtr) (see [Stateless alternatives for server-to-client interactions](#stateless-alternatives-for-server-to-client-interactions)). MRTR is only as available as the unratified `2026-07-28` revision, however, so keep a session if you need server-to-client requests against clients that don't speak it. Note that with `Stateless = false`, a `2026-07-28` request is refused with `UnsupportedProtocolVersion`; the stateful path activates only when a client falls back to an initialize-capable revision.
- **`Stateless = false`** — the right choice when your server depends on sessions for [unsolicited notifications](#how-streamable-http-delivers-messages), resource subscriptions, or per-client isolation, none of which work without a session. Setting this explicitly protects your server from a future default change, and the [MCP specification requires](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports#streamable-http) that clients use sessions when a server's `initialize` response includes an `Mcp-Session-Id` header, so compliant clients always honor your server's session. Server-to-client requests no longer force a session: [elicitation](xref:elicitation) — and the now-deprecated [sampling](xref:sampling) and [roots](xref:roots) — can run statelessly through [MRTR](xref:mrtr) (see [Stateless alternatives for server-to-client interactions](#stateless-alternatives-for-server-to-client-interactions)). Keep a session if you need server-to-client requests against clients that do not support `2026-07-28`. Note that with `Stateless = false`, a `2026-07-28` request is refused with `UnsupportedProtocolVersion`; the stateful path activates only when a client falls back to an initialize-capable revision.
<!-- mlc-disable-next-line -->
> [!TIP]
@ -141,7 +141,7 @@ Most MCP servers fall into this category. Tools that call APIs, query databases,
The traditional approach to server-to-client interactions (elicitation, sampling, roots) requires sessions because the server must hold an open connection to send JSON-RPC requests back to the client. [Multi Round-Trip Requests (MRTR)](xref:mrtr) is a stateless alternative — instead of sending a request, the server returns an **incomplete result** that tells the client what input is needed. The client fulfills the requests and retries the tool call with the responses attached.
This means servers that need user confirmation ([elicitation](xref:elicitation)) — or the now-deprecated (SEP-2577) [sampling](xref:sampling) and [roots](xref:roots) — can run in stateless mode when both sides support MRTR. Because MRTR rides on the unratified `2026-07-28` revision, it is far less broadly supported than session-based requests; keep a session if you must interact with clients that don't speak it.
This means servers that need user confirmation ([elicitation](xref:elicitation)) — or the now-deprecated (SEP-2577) [sampling](xref:sampling) and [roots](xref:roots) — can run in stateless mode when both sides support MRTR. Keep a session if you must interact with clients that do not support `2026-07-28`.
## Stateful mode (sessions)
@ -595,13 +595,13 @@ In stateless mode, each HTTP request creates and disposes a short-lived `McpServ
### Stateless mode
Tasks are a natural fit for stateless servers. The client sends a task-augmented `tools/call` request, receives a task ID immediately, and polls for completion with `tasks/get` or `tasks/result` on subsequent independent HTTP requests. Because each request creates an ephemeral `McpServer` that shares the same `IMcpTaskStore`, all task operations work without any persistent session.
Tasks are a natural fit for stateless servers. The client sends a task-augmented `tools/call` request, receives a task ID immediately, and polls for completion with `tasks/get` on subsequent independent HTTP requests. Because each request creates an ephemeral `McpServer` that shares the same `IMcpTaskStore`, all task operations work without any persistent session.
In stateless mode, there is no `SessionId`, so the task store does not apply session-based isolation. All tasks are accessible from any request to the same server. This is typically fine for single-purpose servers or when authentication middleware already identifies the caller.
### Stateful mode
In stateful mode, the `IMcpTaskStore` receives the session's `SessionId` on every operation: `CreateTaskAsync`, `GetTaskAsync`, `ListTasksAsync`, `CancelTaskAsync`, etc. The built-in <xref:ModelContextProtocol.Extensions.Tasks.InMemoryMcpTaskStore> enforces session isolation: tasks created in one session cannot be accessed from another.
In stateful mode, the `IMcpTaskStore` tracks tasks created for that server instance. The built-in <xref:ModelContextProtocol.Extensions.Tasks.InMemoryMcpTaskStore> keeps tasks in memory; implement a custom store when your scenario requires durable storage or isolation across server instances.
Tasks can outlive individual HTTP requests because the tool executes in the background after returning the initial `CreateTaskResult`. Task cleanup is governed by the task's TTL (time-to-live), not by session termination. However, the `InMemoryMcpTaskStore` loses all tasks if the server process restarts. For durable tasks, implement a custom <xref:ModelContextProtocol.Extensions.Tasks.IMcpTaskStore> backed by an external store. For guidance, see [Implementing a custom task store](xref:tasks#implementing-a-custom-task-store).
@ -653,9 +653,9 @@ When a handler calls `EnablePollingAsync()`:
The `EventStreamStore` itself has TTL-based limits (default: 2-hour event expiration, 30-minute sliding window) that govern event retention, but these do not limit handler concurrency. If you enable `EventStreamStore` on a public-facing server, apply **HTTP rate-limiting middleware** and **reverse proxy limits** to compensate for the loss of stream-level backpressure.
### With tasks (experimental)
### With tasks
[Tasks](xref:tasks) are an experimental feature that enables a "call-now, fetch-later" pattern for long-running tool calls. When a client sends a task-augmented `tools/call` request, the server creates a task record in the <xref:ModelContextProtocol.Extensions.Tasks.IMcpTaskStore>, starts the tool handler as a fire-and-forget background task, and returns the task ID immediately, so the `POST` response completes **before the handler starts its real work**.
[Tasks](xref:tasks) enable a "call-now, fetch-later" pattern for long-running tool calls. When a client sends a task-augmented `tools/call` request, the server creates a task record in the <xref:ModelContextProtocol.Extensions.Tasks.IMcpTaskStore>, starts the tool handler as a fire-and-forget background task, and returns the task ID immediately, so the `POST` response completes **before the handler starts its real work**.
This means: