About MCP Sampling

[GARCHENG]

In the SDK, it only informs us about how to handle sampling requests on the client side. However, there's no guidance on how to send sampling requests from the MCP server. Could you please provide some information on this?

[GARCHENG]

I have found the example for sending sampling request.

[tzolov]

Hey @GARCHENG, the MCP spec is indeed vague about client features like roots and sampling. After working to understand the spec's intention, we've just merged PR #31 that addresses this.

Key point: according to the spec, you can only use listRoots or createMessages within an existing client session. This means these client features are only accessible from within tool/prompt/resource handlers using the new exchange argument.

For a helpful example, check this test:

java-sdk/mcp-spring/mcp-spring-webflux/src/test/java/io/modelcontextprotocol/WebFluxSseIntegrationTests.java

Line 119 in f50f88c

void testCreateMessageSuccess(String clientType) throws InterruptedException {

We're updating the Java SDK documentation, but wanted to share this clarification now.

[GARCHENG]

Hi @tzolov ,

Thank you so much for your #prompt response and assistance. I really appreciate the time you took to address my issue. Your help is greatly appreciated.

Best regards,
GARCHENG

[tzolov]

@GARCHENG,

Latest MCP Java SDK (0.8.1) provides proper sampling and roots support:

• https://modelcontextprotocol.io/sdk/java/mcp-client#sampling-support
• https://modelcontextprotocol.io/sdk/java/mcp-server#using-sampling-from-a-server

Here is a more advanced MCP Sampling example (based on Spring-AI 1.0.0-SNAPSHOT + MCP Java SDK 0.8.1).: https://github.com/spring-projects/spring-ai-examples/tree/main/model-context-protocol/sampling (note that we can retrieve the McpSyncServerExchange from the @Tool's ToolContext

Still missing the related Spring AI MCP (1.0.0-SNAPSHOT) sampling docs. Will come soon

Latest MCP Java SDK (0.8.1) now supports proper sampling and roots:

• Client Sampling Support
• Server-Side Sampling

For a more advanced example using Spring AI 1.0.0-SNAPSHOT + MCP Java SDK 0.8.1, check out:
👉 Spring AI MCP Sampling Example (Note: How You can retrieve McpSyncServerExchange from @Tool's ToolContext.)

📌 Spring AI MCP (1.0.0-SNAPSHOT) sampling docs are still in progress—coming soon!

Let me know if you need more help.

[GARCHENG]

Hi @tzolov ,

Thank you very much for telling me this！ I will study it diligently. I actually studied this part of the content（https://github.com/spring-projects/spring-ai-examples/tree/main/model-context-protocol/sampling） yesterday, and I have a different perspective on the use of mcp sampling. I am very much looking forward to discussing it with you.

After I learning the sampling section of the MCP official documentation ，I believe the purpose of sampling is to clearly inform users of the changes that occur after calling the MCP server tool and to seek the user's consent，similar to how in Cursor, changes to file content require manual acceptance before execution. Therefore, I tried to implement it according to this, and the specific process is as follows：

1. The user requests the client to modify the code of a certain file.
2. After receiving the request, the client requests the LLM.
3. The LLM decides to call a specific tool function.
4. The client then calls the corresponding MPC server.
5. The MPC server sends a sampling request to the client, informing it of the current content of the file, what the modified content will be, and has the LLM organize this information into the appropriate format.
6. The client sends the organized content from the LLM to the user for review.
7. After the user agrees to these changes, the MCP server will actually change the content of the file.

I've put together a simple demo based on these steps, but every time the MCP server processes the sampling response, it times out. After reading through the source code, I found that the timeout duration is hardcoded. If possible, could this timeout duration be made configurable?The corresponding code is as follows: https://github.com/modelcontextprotocol/java-sdk/blob/main/mcp/src/main/java/io/modelcontextprotocol/spec/McpServerSession.java

My ability to express myself is limited, so if I haven't made the above content clear, I sincerely apologize.

Best regards,
GARCHENG

[tzolov]

I believe the purpose of sampling is to clearly inform users of the changes that occur after calling the MCP server tool and to seek the user's consent

I don't thin this assumption is correct. The Sampling Spec clearly states that the purpose of sampling is to "provides a standardized way for servers to request LLM sampling (“completions” or “generations”) from language models via clients."

The Request User Consent functionality, that your are looking for, is not implemented by the Java SDK yet.
It is a SHOULD recommendation that we have not addressed yet.
Please submit a new issue, requesting this feature.

[GARCHENG]

I don't thin this assumption is correct. The Sampling Spec clearly states that the purpose of sampling is to "provides a standardized way for servers to request LLM sampling (“completions” or “generations”) from language models via clients."

Thank you very much for your reply. After re-reading the MCP protocol on sampling, I have deeply realized my previous conceptual misunderstanding of sampling.