MCP Server and Vertex AI

[sidutti]

Bug description
The MCP Server responds the results wrapped as List.
McpSchema.CallToolResult(List.of(new McpSchema.TextContent(callResult)), false);

When using the Vertex AI, the response is marshaled as a Struct (protobuf) which expects a MAP.

In the class VertexAiGeminiChatModel, method messageToGeminiParts when the message is ToolResponseMessage the response string is still being tried to be de-serialized as map. This throws an exception.

Environment
Spring AI Version: 1.0.0-SNAPSHOT
Java : 21

Steps to reproduce
MCP Server : Any ASYNC MCP Server
MCP Client with Vertex AI Chat Model

Expected behavior
The message when is a tool response and if it is an Array, should be mapped and the internal Map should be passed back to VertexAI

[bkowal]

I have been facing the very same issue. I'm also using Vertex AI and MCP client inside Spring AI. The response from MCP server that MCP client inside Spring AI receives seems to be correct. However, when this response is then prepared to be sent to Vertex AI, (VertexAiGeminiChatModel#jsonToStruct(String json)), only a part of the response from the MCP server is actually used. This part of the JSON response is the value of the "content" field. This value is a JSON array. A JSON array doesn't work well with VertexAiGeminiChatModel#jsonToStruct(String json) :-).

java.lang.RuntimeException: com.google.protobuf.InvalidProtocolBufferException: Expect a map object but found: [{"type":"text","text":"[{\"departureAirport\":\"AAA\",\"destinationAirport\":\"BBB\",\"departureDate\":\"2025-05-10\",\"arrivalDate\":\"2025-05-11\"}]"}]
        at org.springframework.ai.vertexai.gemini.VertexAiGeminiChatModel.jsonToStruct(VertexAiGeminiChatModel.java:380) ~[spring-ai-vertex-ai-gemini-1.0.0-M6.jar:1.0.0-M6]
        at org.springframework.ai.vertexai.gemini.VertexAiGeminiChatModel.lambda$messageToGeminiParts$1(VertexAiGeminiChatModel.java:340) ~[spring-ai-vertex-ai-gemini-1.0.0-M6.jar:1.0.0-M6]

I think, the source of this issue is:
org.springframework.ai.mcp.SyncMcpToolCallback#call:

public String call(String functionInput) {
		Map<String, Object> arguments = ModelOptionsUtils.jsonToMap(functionInput);
		CallToolResult response = this.mcpClient
			.callTool(new CallToolRequest(this.getToolDefinition().name(), arguments));
		return ModelOptionsUtils.toJsonString(response.content());
	}

The response looks more or less like this: { "content": [{...}], "isError": false }. Because ModelOptionsUtils.toJsonString(response.content()) uses content(), what is later sent to Vertex AI is [{...}].

As @sidutti wrote:
Expected behavior

The message when is a tool response and if it is an Array, should be mapped and the internal Map should be passed back to VertexAI

OR
org.springframework.ai.mcp.SyncMcpToolCallback#call should leave the original tool response as is (or should leave original response, but drop the "isError" field) instead of extracting the value of the "content" field.

[vikasrangi]

hi Team,

I am also facing similar issue and using gemini 2.5 pro model. Any solutions for this bug in progress?

com.google.protobuf.InvalidProtocolBufferException: Expect a map object but found:

Also, similar bug is reported - #2849

[tzolov]

@sidutti , thank you for reporting this issue.
Let me try to reproduce it.

[markpollack]

please try the snapshots and reopen if it doesn't solve your issue.

[vikasrangi]

It's working. Thanks a lot.