[Server] Allow tools to return structured content without text content #521

Is your feature request related to a problem? Please describe.

When a tool returns structured data, the SDK currently produces both:

• a serialized JSON TextContent representation in content;
• the same data in structuredContent.

For clients that consume structuredContent, the text representation is redundant and can significantly increase the size of the tool response. In some integrations, large duplicated JSON content may also be stored, truncated, or paginated unnecessarily.

This was discussed in symfony/ai#2200.

The current workaround is for every tool to inject RequestContext and manually construct a CallToolResult, deciding whether to include content depending on the client.

This also couples individual tools to client-specific behavior.

Describe the solution you'd like

Allow a tool to explicitly opt into structured-content-only results.

For example, an attribute-level option:

#[McpTool(structuredContentOnly: true)]
public function getData(): array
{
    return [
        'foo' => 'bar',
    ];
}

When enabled, the SDK would produce a CallToolResult with:

new CallToolResult(
    content: [],
    structuredContent: [
        'foo' => 'bar',
    ],
);

The default behavior should remain unchanged.

Ideally, the option should be handled inside the SDK's existing tool result formatting pipeline, rather than requiring individual tools to construct CallToolResult themselves.

This would provide an explicit per-tool opt-in without introducing client-specific logic into the SDK.

Describe alternatives you've considered

1. Manually construct CallToolResult in every tool

   A tool can currently return a CallToolResult directly and decide whether to include content or structuredContent.

   However, this requires application-level code to know about the response-shaping details and duplicates the same logic across tools.

2. Use RequestContext and inspect the connected client

   A tool can inspect the client information through RequestContext and choose the response shape dynamically.

   This works as a workaround, but it couples every tool to client detection and requires repetitive boilerplate.

3. Automatically detect the client and choose the response format

   This could be useful as a future enhancement, but it seems like a separate concern from the basic SDK capability.

   A per-tool opt-in would provide a small, predictable first step while keeping automatic client-aware shaping as a possible follow-up.

Additional context

The Symfony AI discussion identified that this behavior belongs in mcp/sdk, rather than symfony/mcp-bundle, because the actual result shaping is implemented by the SDK.

Relevant SDK code includes:

• ToolReference::formatResult()
• ToolReference::extractStructuredContent()
• the tool call handler that builds CallToolResult

See the discussion in symfony/ai#2200 and the follow-up documentation PR #2506.

The MCP protocol already models content and structuredContent as separate parts of CallToolResult; content is an array, so an empty content array can represent a structured-content-only result.

The SDK could preserve the current behavior by default and only omit the generated TextContent when the new option is explicitly enabled.