chore(dev): properly process MCP results

backend/agent/run.py

@@ -170,6 +170,20 @@ async def run_agent(
             qualified_name = mcp_config.get('qualifiedName', 'unknown')
             mcp_info += f"- {server_name} (use prefix: mcp_{qualified_name}_)\n"
         
+        # Add critical instructions for using search results
+        mcp_info += "\n🚨 CRITICAL MCP TOOL RESULT INSTRUCTIONS 🚨\n"
+        mcp_info += "When you use ANY MCP (Model Context Protocol) tools:\n"
+        mcp_info += "1. ALWAYS read and use the EXACT results returned by the MCP tool\n"
+        mcp_info += "2. For search tools: ONLY cite URLs, sources, and information from the actual search results\n"
+        mcp_info += "3. For any tool: Base your response entirely on the tool's output - do NOT add external information\n"
+        mcp_info += "4. DO NOT fabricate, invent, hallucinate, or make up any sources, URLs, or data\n"
+        mcp_info += "5. If you need more information, call the MCP tool again with different parameters\n"
+        mcp_info += "6. When writing reports/summaries: Reference ONLY the data from MCP tool results\n"
+        mcp_info += "7. If the MCP tool doesn't return enough information, explicitly state this limitation\n"
+        mcp_info += "8. Always double-check that every fact, URL, and reference comes from the MCP tool output\n"
+        mcp_info += "\nIMPORTANT: MCP tool results are your PRIMARY and ONLY source of truth for external data!\n"
+        mcp_info += "NEVER supplement MCP results with your training data or make assumptions beyond what the tools provide.\n"
+        
         system_content += mcp_info
     
     system_message = { "role": "system", "content": system_content }

backend/agent/tools/mcp_tool_wrapper.py

@@ -134,26 +134,41 @@ class MCPToolWrapper(Tool):
             server_name = parts[1] if len(parts) > 1 else "unknown"
             original_tool_name = parts[2] if len(parts) > 2 else tool_name
             
-            # Enhance the result with metadata for better frontend display
-            enhanced_result = {
-                "mcp_metadata": {
-                    "server_name": server_name,
-                    "tool_name": original_tool_name,
-                    "full_tool_name": tool_name,
-                    "arguments_count": len(arguments) if isinstance(arguments, dict) else 0,
-                    "is_mcp_tool": True
-                },
-                "content": result.get("content", ""),
-                "isError": result.get("isError", False),
-                "raw_result": result
-            }
-            
             # Check if it's an error
             if result.get("isError", False):
-                return self.fail_response(json.dumps(enhanced_result, indent=2))
+                error_result = {
+                    "mcp_metadata": {
+                        "server_name": server_name,
+                        "tool_name": original_tool_name,
+                        "full_tool_name": tool_name,
+                        "arguments_count": len(arguments) if isinstance(arguments, dict) else 0,
+                        "is_mcp_tool": True
+                    },
+                    "content": result.get("content", ""),
+                    "isError": True,
+                    "raw_result": result
+                }
+                return self.fail_response(json.dumps(error_result, indent=2))
+            
+            # Format the result in an LLM-friendly way with content first
+            actual_content = result.get("content", "")
+            
+            # Create a clear, LLM-friendly response that puts the content first
+            llm_friendly_result = f"""MCP Tool Result from {server_name.upper()}:
+
+{actual_content}
+
+---
+Tool Metadata: {json.dumps({
+    "server": server_name,
+    "tool": original_tool_name,
+    "full_tool_name": tool_name,
+    "arguments_used": arguments,
+    "is_mcp_tool": True
+}, indent=2)}"""
                 
-            # Return successful result with enhanced metadata
-            return self.success_response(json.dumps(enhanced_result, indent=2))
+            # Return successful result with LLM-friendly formatting
+            return self.success_response(llm_friendly_result)
             
         except ValueError as e:
             # Handle specific MCP errors (like invalid tool name format)

backend/agentpress/response_processor.py

@@ -1456,6 +1456,36 @@ class ResponseProcessor:
                 )
                 return message_obj # Return the full message object
             
+            # Check if this is an MCP tool (function_name starts with "call_mcp_tool")
+            function_name = tool_call.get("function_name", "")
+            if function_name == "call_mcp_tool":
+                # Special handling for MCP tools - make content prominent and LLM-friendly
+                result_role = "user" if strategy == "user_message" else "assistant"
+                
+                # Extract the actual content from the ToolResult
+                if hasattr(result, 'output'):
+                    mcp_content = str(result.output)
+                else:
+                    mcp_content = str(result)
+                
+                # Create a simple, LLM-friendly message format that puts content first
+                simple_message = {
+                    "role": result_role,
+                    "content": mcp_content  # Direct content, no complex nesting
+                }
+                
+                logger.info(f"Adding MCP tool result with simplified format for LLM visibility")
+                self.trace.event(name="adding_mcp_tool_result_simplified", level="DEFAULT", status_message="Adding MCP tool result with simplified format for LLM visibility")
+                
+                message_obj = await self.add_message(
+                    thread_id=thread_id, 
+                    type="tool",
+                    content=simple_message,
+                    is_llm_message=True,
+                    metadata=metadata
+                )
+                return message_obj
+            
             # For XML and other non-native tools, use the new structured format
             # Determine message role based on strategy
             result_role = "user" if strategy == "user_message" else "assistant"

frontend/src/components/sidebar/sidebar-left.tsx

@@ -156,7 +156,7 @@ export function SidebarLeft({
             <Store className="h-4 w-4" />
             <span className="flex items-center justify-between w-full">
               Marketplace
-              <Badge variant="secondary" className="ml-2 text-xs border bg-green-200/50 dark:bg-green-900/40 text-green-700 dark:text-green-300">
+              <Badge variant="secondary" className="ml-2 text-xs border bg-blue-200/50 dark:bg-blue-900/40 text-blue-700 dark:text-blue-300">
                 New
               </Badge>
             </span>