feat: Add MCP structuredContent support (fixes #283)

[jeff52415]

Implement MCP structuredContent support to resolve data loss issue

This branch addresses Issue #283 by implementing proper support for MCP's
structuredContent field in CallToolResult responses. Previously, the
langchain-mcp-adapters library ignored the structuredContent
field.

Key Changes:
• Refactored _convert_call_tool_result() to handle structuredContent
• Updated return format to use LangChain's content_and_artifact pattern

The implementation ensures:

• structuredContent is preserved in artifact["structuredContent"]
• Backward compatibility with existing text/non-text content handling
• Compliance with MCP specification (2025-06-18)

Example:

from langchain_core.messages import ToolCall
result = await tool.ainvoke(
    ToolCall(
        name="tool_name",
        args={....},
        id="call_123",  # any unique string ID
        type="tool_call",
    )
    
)
print(result.content)
if result.artifact:
    print(result.artifact.get("structuredContent", None))
    print(result.artifact.get("nonText", None))

Fixes: #283

[IoannisMaras]

Can confirm this works . And I need it as well

[eyurtsev]

Related: https://github.com/langchain-ai/langchain-mcp-adapters/pull/107/files

[eyurtsev]

How is your application consuming this information? Do you have an agent or is it a fixed workflow where you access the results of the tool call? If it's an agent, do you loop through the message history and check the artifacts field?

[jeff52415]

@eyurtsev Thanks for asking.
content_and_artifact pattern only propagates through the full flow when emitted via the ToolCall interface.
This is how I use to integrate structured outputs into the agent pipeline.

async def tool_node(state: AgentState) -> dict[str, list[ToolMessage]]:
    """Execute tool calls by constructing ToolCall objects for each call."""
    last = state["messages"][-1]
    outputs: list[ToolMessage] = []
    for call in getattr(last, "tool_calls", []) or []:
        # Build ToolCall to satisfy adapters requiring ToolCall input.
        name = call["name"]
        args = call["args"]
        if call.get("id") is None:
            call_id = str(uuid.uuid4())
        else:
            call_id = call["id"]
        tool_call = ToolCall(name=name, args=args, id=call_id, type="tool_call")

        tool = tools_by_name[name]
        if hasattr(tool, "ainvoke"):
            result = await tool.ainvoke(tool_call)  # type: ignore[attr-defined]
        else:
            # Fall back to sync execution if async not available
            result = tool.invoke(tool_call)
    output = "some post-processing......."
    return {"messages": outputs}

workflow = StateGraph(AgentState)
workflow.add_node("tools", tool_node)