modelcontextprotocol/csharp-sdk
Watch
1
0
Fork
You've already forked csharp-sdk
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Multi Round-Trip Requests (MRTR) (#1458)

Browse source
This commit is contained in:
Stephen Halter 2026-06-03 16:19:31 -07:00 • committed by GitHub
commit f6dbe4372f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 5832 additions and 103 deletions
Download patch file Download diff file Expand all files Collapse all files

291
docs/concepts/mrtr/mrtr.md Normal file
View file

@ -0,0 +1,291 @@
---
title: Multi Round-Trip Requests (MRTR)
author: halter73
description: How servers request client input during tool execution using Multi Round-Trip Requests.
uid: mrtr
---
# Multi Round-Trip Requests (MRTR)
<!-- mlc-disable-next-line -->
> [!WARNING]
> MRTR is part of the **`DRAFT-2026-v1`** revision of the MCP specification ([SEP-2322](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2322)). The wire format and API surface may change before the revision is ratified. See the [Experimental APIs](../../experimental.md) documentation for details on working with experimental APIs.
Multi Round-Trip Requests (MRTR) let a server tool request input from the client — such as [elicitation](xref:elicitation), [sampling](xref:sampling), or [roots](xref:roots) — as part of a single tool call, without requiring a separate server-to-client JSON-RPC request for each interaction. Instead of returning a final result, the server returns an **incomplete result** containing one or more input requests. The client fulfills those requests and retries the original tool call with the responses attached.
## Overview
MRTR is useful when:
- A tool needs user confirmation before proceeding (elicitation).
- A tool needs LLM reasoning from the client (sampling).
- A tool needs an updated list of client roots.
- A tool needs to perform multiple rounds of interaction in a single logical operation.
- A stateless server needs to orchestrate multi-step flows without keeping handler state in memory between rounds.
## How MRTR works
1. The client calls a tool on the server via `tools/call`.
2. The server tool determines it needs client input and returns an `InputRequiredResult` containing `inputRequests` and/or `requestState`.
3. The client resolves each input request (for example by prompting the user for elicitation, calling an LLM for sampling, or listing its roots).
4. The client retries the original `tools/call` with `inputResponses` (keyed to the input requests) and `requestState` echoed back.
5. The server processes the responses and either returns a final result or another `InputRequiredResult` for additional rounds.
## Opting in
MRTR activates when both peers negotiate protocol revision **`DRAFT-2026-v1`** during `initialize`. The C# SDK opts in by listing `DRAFT-2026-v1` as a supported protocol version on the client; servers automatically accept it when offered. No experimental flags are required.
```csharp
// Client
var clientOptions = new McpClientOptions
{
ProtocolVersion = "DRAFT-2026-v1",
Handlers = new McpClientHandlers
{
ElicitationHandler = HandleElicitationAsync,
SamplingHandler = HandleSamplingAsync,
}
};
```
Under `DRAFT-2026-v1`, MRTR is the recommended way to obtain client input from a server handler. The spec removes the legacy server-to-client `elicitation/create`, `sampling/createMessage`, and `roots/list` request methods, so any code that needs to work on a `DRAFT-2026-v1` Streamable HTTP server (which will be stateless-only in a future revision) must use `InputRequiredException` rather than <xref:ModelContextProtocol.Server.McpServer.ElicitAsync*>, <xref:ModelContextProtocol.Server.McpServer.SampleAsync*>, or <xref:ModelContextProtocol.Server.McpServer.RequestRootsAsync*>. The legacy methods still work on stateful sessions — that's how stdio servers keep working under draft today — but they throw `InvalidOperationException("X is not supported in stateless mode.")` on any stateless session, current or draft.
Under the current protocol revision (`2025-06-18` and earlier), `InputRequiredException` is still supported in stateful sessions via a backward-compatibility resolver — see [Compatibility](#compatibility) below.
## Authoring an MRTR tool
A tool participates in MRTR by throwing <xref:ModelContextProtocol.Protocol.InputRequiredException> with an <xref:ModelContextProtocol.Protocol.InputRequiredResult> describing what it needs. On retry, the client's responses arrive on the request parameters and the tool inspects them to decide what to do next.
### Checking MRTR support
Tools should check <xref:ModelContextProtocol.Server.McpServer.IsMrtrSupported> before throwing `InputRequiredException`. It returns `true` when either:
- The negotiated protocol revision is `DRAFT-2026-v1` (MRTR is native), or
- The session is stateful under the current protocol (the SDK can resolve input requests via legacy JSON-RPC and retry the handler).
```csharp
[McpServerTool, Description("A tool that uses MRTR")]
public static string MyTool(
McpServer server,
RequestContext<CallToolRequestParams> context)
{
if (!server.IsMrtrSupported)
{
return "This tool requires a client that negotiates DRAFT-2026-v1, "
+ "or a stateful current-protocol session.";
}
// ... MRTR logic
}
```
### Returning an incomplete result
Throw <xref:ModelContextProtocol.Protocol.InputRequiredException> to return an incomplete result. The exception carries an <xref:ModelContextProtocol.Protocol.InputRequiredResult> containing `inputRequests` and/or `requestState`:
```csharp
[McpServerTool, Description("Tool managing its own MRTR flow")]
public static string AnswerTool(
McpServer server,
RequestContext<CallToolRequestParams> context,
[Description("The user's question")] string question)
{
var requestState = context.Params!.RequestState;
var inputResponses = context.Params!.InputResponses;
// On retry, process the client's responses
if (requestState is not null && inputResponses is not null)
{
var elicitResult = inputResponses["user_answer"].Deserialize(InputResponse.ElicitResultJsonTypeInfo);
return $"You answered: {elicitResult?.Content?.FirstOrDefault().Value}";
}
if (!server.IsMrtrSupported)
{
return "MRTR is not supported by this client.";
}
// First call — request user input
throw new InputRequiredException(
inputRequests: new Dictionary<string, InputRequest>
{
["user_answer"] = InputRequest.ForElicitation(new ElicitRequestParams
{
Message = $"Please answer: {question}",
RequestedSchema = new()
{
Properties = new Dictionary<string, ElicitRequestParams.PrimitiveSchemaDefinition>
{
["answer"] = new ElicitRequestParams.StringSchema
{
Description = "Your answer"
}
}
}
})
},
requestState: "awaiting-answer");
}
```
### Accessing retry data
When the client retries a tool call, the retry data is available on the request parameters:
- <xref:ModelContextProtocol.Protocol.RequestParams.InputResponses> — a dictionary of client responses keyed by the same keys used in `inputRequests`.
- <xref:ModelContextProtocol.Protocol.RequestParams.RequestState> — the opaque state string echoed back by the client.
Use <xref:ModelContextProtocol.Protocol.InputResponse.Deserialize*> with the `JsonTypeInfo<T>` matching the response type. The expected type follows from the matching <xref:ModelContextProtocol.Protocol.InputRequest.Method> in the original `inputRequests` map — there is no on-the-wire discriminator.
- Elicitation — `response.Deserialize(InputResponse.ElicitResultJsonTypeInfo)`
- Sampling — `response.Deserialize(InputResponse.CreateMessageResultJsonTypeInfo)`
- Roots list — `response.Deserialize(InputResponse.ListRootsResultJsonTypeInfo)`
### Load shedding with requestState-only responses
A server can return a `requestState`-only incomplete result (without any `inputRequests`) to defer processing. This is useful for load shedding or breaking up long-running work across multiple requests:
```csharp
[McpServerTool, Description("Tool that defers work using requestState")]
public static string DeferredTool(
McpServer server,
RequestContext<CallToolRequestParams> context)
{
var requestState = context.Params!.RequestState;
if (requestState is not null)
{
// Resume deferred work
var state = JsonSerializer.Deserialize<MyState>(
Convert.FromBase64String(requestState));
return $"Completed step {state!.Step}";
}
if (!server.IsMrtrSupported)
{
return "MRTR is not supported by this client.";
}
// Defer work to a later retry
var initialState = new MyState { Step = 1 };
throw new InputRequiredException(
requestState: Convert.ToBase64String(
JsonSerializer.SerializeToUtf8Bytes(initialState)));
}
```
The client automatically retries `requestState`-only incomplete results, echoing the state back without needing to resolve any input requests.
### Multiple round trips
A tool can perform multiple rounds of interaction by throwing `InputRequiredException` multiple times across retries. Use `requestState` to track which round you're on:
```csharp
[McpServerTool, Description("Multi-step wizard")]
public static string WizardTool(
McpServer server,
RequestContext<CallToolRequestParams> context)
{
var requestState = context.Params!.RequestState;
var inputResponses = context.Params!.InputResponses;
if (requestState == "step-2" && inputResponses is not null)
{
var name = inputResponses["name"].Deserialize(InputResponse.ElicitResultJsonTypeInfo)?.Content?.FirstOrDefault().Value;
var age = inputResponses["age"].Deserialize(InputResponse.ElicitResultJsonTypeInfo)?.Content?.FirstOrDefault().Value;
return $"Welcome, {name}! You are {age} years old.";
}
if (requestState == "step-1" && inputResponses is not null)
{
var name = inputResponses["name"].Deserialize(InputResponse.ElicitResultJsonTypeInfo)?.Content?.FirstOrDefault().Value;
// Second round — ask for age
throw new InputRequiredException(
inputRequests: new Dictionary<string, InputRequest>
{
["age"] = InputRequest.ForElicitation(new ElicitRequestParams
{
Message = $"Hi {name}! How old are you?",
RequestedSchema = new()
{
Properties = new Dictionary<string, ElicitRequestParams.PrimitiveSchemaDefinition>
{
["age"] = new ElicitRequestParams.NumberSchema
{
Description = "Your age"
}
}
}
})
},
requestState: "step-2");
}
if (!server.IsMrtrSupported)
{
return "MRTR is not supported. Please use a compatible client.";
}
// First round — ask for name
throw new InputRequiredException(
inputRequests: new Dictionary<string, InputRequest>
{
["name"] = InputRequest.ForElicitation(new ElicitRequestParams
{
Message = "What's your name?",
RequestedSchema = new()
{
Properties = new Dictionary<string, ElicitRequestParams.PrimitiveSchemaDefinition>
{
["name"] = new ElicitRequestParams.StringSchema
{
Description = "Your name"
}
}
}
})
},
requestState: "step-1");
}
```
### Providing custom error messages
When MRTR is not supported, you can provide domain-specific guidance:
```csharp
if (!server.IsMrtrSupported)
{
return "This tool requires interactive input. To use it:\n"
+ "1. Connect with a client that negotiates MCP protocol revision DRAFT-2026-v1, or\n"
+ "2. Use a stateful current-protocol session so the server can resolve the input requests for you.\n"
+ "\nStateless current-protocol sessions cannot resolve MRTR input requests.";
}
```
## Compatibility
The SDK supports `InputRequiredException` across two protocol revisions and two session modes:
| Negotiated protocol | Session mode | Behavior |
|---|---|---|
| `DRAFT-2026-v1` | Stateful | Native MRTR — `InputRequiredResult` is serialized directly to the wire. |
| `DRAFT-2026-v1` | Stateless | Native MRTR — `InputRequiredResult` is serialized directly to the wire. No server-side handler state needed. |
| Current (`2025-06-18` and earlier) | Stateful | Backward-compatibility resolver — the SDK sends standard `elicitation/create` / `sampling/createMessage` / `roots/list` JSON-RPC requests to the client, collects the responses, and retries the handler with `inputResponses` populated. Up to 10 retry rounds. |
| Current (`2025-06-18` and earlier) | Stateless | **Not supported** — `InputRequiredException` raises an `McpException`. The client doesn't speak MRTR, and the server can't resolve input requests via JSON-RPC without a persistent session. |
> [!NOTE]
> The backcompat resolver is intentionally limited to 10 retry rounds. Tools that need more rounds should require `DRAFT-2026-v1` (check `IsMrtrSupported`).
### Why `ElicitAsync` / `SampleAsync` / `RequestRootsAsync` throw on stateless servers
`ElicitAsync` / `SampleAsync` / `RequestRootsAsync` issue a JSON-RPC request to the client and wait for the response on the same session. Stateless servers don't have a persistent session to wait on, so the SDK fails fast with `InvalidOperationException("X is not supported in stateless mode.")` (the check is `McpServer.ClientCapabilities is null`, which is the SDK's proxy for stateless).
Under the current protocol revision (`2025-06-18` and earlier), stdio and stateful Streamable HTTP keep `ClientCapabilities` populated, so the legacy methods work normally and remain the recommended way to do one-shot client interactions. Under `DRAFT-2026-v1`, the spec removes those request methods from Streamable HTTP entirely; the SDK still allows the legacy methods on draft stdio sessions because stdio is implicitly single-process / stateful and the client handler is wired up regardless of negotiated revision. `InputRequiredException` is the way to write tools that work on every supported configuration.
### Future direction
The `DRAFT-2026-v1` revision is moving toward a stateless-only model: `Mcp-Session-Id` is being removed, and Streamable HTTP servers will run statelessly by default under the draft revision. When that lands, the `Stateful` row for `DRAFT-2026-v1` in the compatibility matrix above collapses into the `Stateless` row (Streamable HTTP under draft becomes stateless-only), and `InputRequiredException` becomes uniformly required for non-stdio servers. The current-protocol resolver path will remain for backward compatibility with older clients and stateful servers.
This work is a follow-up to the present PR.