SyncMcpToolCallback call fails with UnsupportedOperationException if ToolContext is set

[henningSaulCM]

Bug description
Integrating and calling an MCP Tool with an existing Spring AI application that uses a ToolContext will result in an UnsupportedOperationException.

Environment
Spring AI 1.0.0-M6

Steps to reproduce

1. Integrate https://github.com/modelcontextprotocol/servers/tree/main/src/everything into a Spring AI application using https://docs.spring.io/spring-ai/reference/api/mcp/mcp-client-boot-starter-docs.html
2. Set the ToolContext in your Spring AI application
3. Call one of the tools provided by the everything MCP server, e.g. printEnv
4. Observe "UnsupportedOperationException: Tool context is not supported!" for SyncMcpToolCallback

Exception is raised here:

spring-ai/spring-ai-core/src/main/java/org/springframework/ai/tool/ToolCallback.java

Line 57 in 2394ac8

throw new UnsupportedOperationException("Tool context is not supported!");

Expected behavior
The MCP printEnv tool is called, ignoring the ToolContext (assuming this concept is not supported in MCP).

[tzolov]

Thanks @henningSaulCM , I believe this is a know issue and should have been resolved by
the 82d98b6
Can you test with 1.0.0-SNAPSHOT please?

[tzolov]

Likely the same as #2337

[henningSaulCM]

Hi @tzolov, thanks for chiming in.

I just tried with 1.0.0-SNAPSHOT (spring-ai-core-1.0.0-20250307.123602-1325.jar) and it seems I'm still getting the UnsupportedOperationException.

Stacktrace (excerpt):

java.lang.UnsupportedOperationException: Tool context is not supported!
		at org.springframework.ai.tool.ToolCallback.call(ToolCallback.java:57)
		at org.springframework.ai.model.tool.DefaultToolCallingManager.executeToolCall(DefaultToolCallingManager.java:226)
		at org.springframework.ai.model.tool.DefaultToolCallingManager.executeToolCalls(DefaultToolCallingManager.java:138)
		at org.springframework.ai.openai.OpenAiChatModel.lambda$internalStream$10(OpenAiChatModel.java:442)
		at reactor.core.publisher.FluxFlatMap$FlatMapMain.onNext(FluxFlatMap.java:388)
		...

Any thoughts?

[tzolov]

@henningSaulCM I overlooked that AsyncMcpToolCallback and SyncMcpToolCallback don't implement String call(String toolInput, ToolContext tooContext).
I'm not sure how to map ToolContext to a general-purpose MCP Tool Call, as the context only makes sense within Spring AI Tool Calling.
Are you looking for MCP ToolContext support, or is the error happening because you're mixing MCP and Spring AI tools? If it's the latter, we could implement the method for both callback types by ignoring the toolContext. Would that work?

[henningSaulCM]

Hi @tzolov,

Are you looking for MCP ToolContext support, or is the error happening because you're mixing MCP and Spring AI tools? If it's the latter, we could implement the method for both callback types by ignoring the toolContext. Would that work?

Yes, this is happening because I was trying to integrate MCP tools with an existing Spring AI application that uses custom Spring AI tools and custom (non-empty) ToolContext. If MCP does not support a ToolContext (yet), ignoring the ToolContext for MCP Tool calls would be just fine IMHO to avoid the Exception.

#2432 looks good to me.
Thanks for your work on Spring AI and the quick turnaround for this particular issue.

Cheers

[gule002]

@tzolov when you said " as the context only makes sense within Spring AI Tool Calling." can you please elaborate. Do you mean this is supported only when the tool is not remotely hosted over sse?

I am attempting to pass a bearer token back to the MCP server using some mechanism - I thought ToolContext would be the right way, but if it is not supported, I am at a loss on how to pass a special parameter back that the AI agent is not allowed to touch.

[ligenzatomas]

@tzolov when you said " as the context only makes sense within Spring AI Tool Calling." can you please elaborate. Do you mean this is supported only when the tool is not remotely hosted over sse?

I am attempting to pass a bearer token back to the MCP server using some mechanism - I thought ToolContext would be the right way, but if it is not supported, I am at a loss on how to pass a special parameter back that the AI agent is not allowed to touch.

I am also interested in this topic.

[gule002]

I think I got to the bottom of this, currently the framework and maybe even the protocol (SSE/STDIO transport) does not support the ToolContext feature, I got it working with great success by pulling the tool into the App server's memory and process space by using org.springframework.ai.tool.method.MethodToolCallback

In order to get a reference to the MethodToolCallback objects you annotated and scanned in your App server, just autowire it:

List<ToolCallbackProvider> toolCallbackProviders;

This allows you to pass in the HttpServletRequest object and do all sorts of fancy things:

            responseSpec = chatClient.prompt(prompt)
                .toolContext(Map.of("httpRequest", httpRequest))
                .tools(selectedTools.toArray(new ToolCallback[0]))
                .call();

For example, you can have the tool return the time in the visitor's own timezone:

        @Tool(description = "Format a date string according to the specified pattern in the user's timezone")
        public String formatDate(
            @ToolParam(description = "ISO date-time string to format (e.g., '2023-01-01T12:00:00Z')") String dateString,
            @ToolParam(description = "Date format pattern (e.g., 'yyyy-MM-dd', 'MMMM d, yyyy', etc.)") String pattern,
            ToolContext context
        ) {
            try {
                HttpServletRequest request = (HttpServletRequest) context.getContext().get("httpRequest");
                ZoneId userZone = getUserTimeZone(request);
                
                // Parse the input date
                Instant instant = Instant.parse(dateString);
                ZonedDateTime dateTime = instant.atZone(userZone);
                
                // Format using the provided pattern
                DateTimeFormatter formatter = DateTimeFormatter.ofPattern(pattern);
                return dateTime.format(formatter);
            } catch (Exception e) {
                logger.error("Error formatting date: {}", e.getMessage());
                return "Error formatting date: " + e.getMessage();
            }
        }