diff --git a/src/google/adk/models/lite_llm.py b/src/google/adk/models/lite_llm.py
index 9e3698b190..f47485e3c2 100644
--- a/src/google/adk/models/lite_llm.py
+++ b/src/google/adk/models/lite_llm.py
@@ -367,11 +367,21 @@ def _content_to_message_param(
   tool_messages = []
   for part in content.parts:
     if part.function_response:
+      # FIX: Check if response is already a string before serializing.
+      # MCP tool responses come as JSON strings, but _safe_json_serialize was
+      # double-serializing them (json.dumps on already-JSON strings), causing
+      # triple-nested JSON like: '{"content": [{"type": "text", "text": "{\n \"type\"..."}]}'
+      # This prevented Claude/GPT from parsing tool results correctly.
+      response_content = (
+          part.function_response.response
+          if isinstance(part.function_response.response, str)
+          else _safe_json_serialize(part.function_response.response)
+      )
       tool_messages.append(
           ChatCompletionToolMessage(
               role="tool",
               tool_call_id=part.function_response.id,
-              content=_safe_json_serialize(part.function_response.response),
+              content=response_content,
           )
       )
   if tool_messages: