Support Langchain MultiServerMCPClient

[lemaiwo]

Describe the Problem

Dear, Would it be possible to support LangChain MultiServerMCPClient in combination with SAP AI SDK? Langchain integration is supported but not the mcp server client, see documentation here: https://sap.github.io/ai-sdk/docs/js/langchain/orchestration

Propose a Solution

Support MCP Server client for langchain in SAP AI SDK

Describe Alternatives

No response

Affected Development Phase

Production

Impact

Blocked

Timeline

No response

Additional Context

No response

[ZhongpinWang]

Hi @lemaiwo ,

Thanks for reaching out. Actually, both our Azure OpenAI and Orchestration clients already work with LangChain's MCP adapter. 😉

Have a look at the documentation -> https://sap.github.io/ai-sdk/docs/js/tutorials/getting-started-with-agents.

The same tutorial is also available on the sample code

ai-sdk-js/sample-code/src/tutorials/mcp/mcp-adapter.ts

Line 6 in 12286c0

const client = new MultiServerMCPClient({

If there is any more question, feel free to re-open.

[lemaiwo]

Do you have an example without the workflow? Would it be possible to just use it like this with the invoke function or is a workflow required?

`

        const mcpClient = new MultiServerMCPClient({
            useStandardContentBlocks: true,
            throwOnLoadError: true,
            mcpServers: {
                logServer: {
                    url: serverUrl + '/mcp',
                    transport: "http" as const,
                    reconnect: {
                        enabled: true,
                        maxAttempts: 3,
                        delayMs: 1000
                    }
                }
            }
        });

        // Get tools from MCP servers
        const mcpTools = await mcpClient.getTools();
        const model = new LangChainOClient({
            promptTemplating: {
                model: {
                    name: 'gpt-4o-mini',
                    params: {
                        temperature: 0.7
                    }
                }
            }
        });
        // create a model with access to the tools
        const modelWithTools = model.bindTools(mcpTools);
        const response = await modelWithTools.invoke("Can you analyze the log and provide a summary of all the logs?");
        return response;`

I tried this but the response was empty.

[lemaiwo]

This is the response we always get:

It doesn't seem to work with langchain in general

[lemaiwo]

The same mcp with langchain and anthropic llm does work not using SAP AI SDK

[lemaiwo]

When I do use StateGraph I get the following error:

code = 400
location = 'Request Body'
message = "400 - Request Body: Additional properties are not allowed ('tool_call_id' was unexpected)"

[lemaiwo]

It does seem to work when using AzureOpenAiChatClient

[ZhongpinWang]

Hi @lemaiwo ,

I have investigated into this issue and here are my feedback.

First of all, we are not doing much when it comes to MCP support as @langchain/mcp-adapters comes from LangChain and we only need to guarantee our LangChain Orchestration client from @sap-ai-sdk/langchain correctly implementing LangChain's bindTools() interface. The rest is handled by LangChain / LangGraph. So I would be surprised if it works in LangGraph but not directly with LangChain since in LangGraph, we also need to call bindTools() from LangChain.

There was one issue found in our tutorial (which resides in our dummy mcp weather server side). We need to import * as z from 'zod/v3'; instead of zod/v4 from zod@3.x package since @modelcontextprotocol/sdk's zod dependency version is totally outdated, which causes incompatibility issue with the latest zod@3.x used in our SDK (and in LangChain). Both LangChain and @modelcontextprotocol/sdk are planning to fully support zod@4.x package.

But this finding might be irrelevant to your question, as in your previous comment, you complained about an "empty response" you received. #1145 (comment)
The response is not really empty, as you can see in your screenshot, the finish reason is tool_calls and you need to further call the tool, send back a tool message to LLM and then finally get the final response. This tool calling part is fully automated by LangGraph but not in LangChain or OpenAI, this is also why LangGraph is more advanced and more suitable for agent use case and LangChain only send individual LLM calls. I recommend reading https://platform.openai.com/docs/guides/function-calling#the-tool-calling-flow which is a quite neat explanation and this is the same process in LangChain (and automated by LangGraph).

Nevertheless, I created a small sample for using LangChain's MCP adapter with pure LangChain (together with the aforementioned fix) in #1162.

Feel free to provide your feedback.