diff --git a/backend/app/api/endpoints/assistant_chat.py b/backend/app/api/endpoints/assistant_chat.py
index 83422367..3bcc1f5e 100644
--- a/backend/app/api/endpoints/assistant_chat.py
+++ b/backend/app/api/endpoints/assistant_chat.py
@@ -145,11 +145,17 @@ async def post_message(
 
     plan = await get_user_plan(current_user.account_id, db)
 
+    # Capture scalar fields before the try block — after db.rollback()
+    # the ORM objects are expired and accessing attributes triggers a
+    # lazy load, which crashes in async context (MissingGreenlet).
+    user_id = current_user.id
+    account_id = current_user.account_id
+
     try:
         ai_content, suggested_flows, chat = await assistant_chat_service.send_message(
             chat_id=chat_id,
-            user_id=current_user.id,
-            account_id=current_user.account_id,
+            user_id=user_id,
+            account_id=account_id,
             message=data.message,
             db=db,
         )
@@ -159,8 +165,8 @@ async def post_message(
         logger.exception("Assistant chat message failed: %s", e)
         await db.rollback()
         await record_ai_usage(
-            user_id=current_user.id,
-            account_id=current_user.account_id,
+            user_id=user_id,
+            account_id=account_id,
             conversation_id=None,
             generation_type="assistant_message",
             tier=plan,
@@ -180,8 +186,8 @@ async def post_message(
         )
 
     await record_ai_usage(
-        user_id=current_user.id,
-        account_id=current_user.account_id,
+        user_id=user_id,
+        account_id=account_id,
         conversation_id=None,
         generation_type="assistant_message",
         tier=plan,
diff --git a/backend/app/services/assistant_chat_service.py b/backend/app/services/assistant_chat_service.py
index 86390608..9508fe56 100644
--- a/backend/app/services/assistant_chat_service.py
+++ b/backend/app/services/assistant_chat_service.py
@@ -189,15 +189,29 @@ async def _call_anthropic_cached(
             }
         ]
 
-    response = await client.beta.messages.create(
-        model=settings.AI_MODEL_ANTHROPIC,
-        max_tokens=max_tokens,
-        system=system_blocks,
-        messages=messages,
-        mcp_servers=mcp_servers,
-        tools=tools,
-        betas=["mcp-client-2025-11-20"],
-    )
+    try:
+        response = await client.beta.messages.create(
+            model=settings.AI_MODEL_ANTHROPIC,
+            max_tokens=max_tokens,
+            system=system_blocks,
+            messages=messages,
+            mcp_servers=mcp_servers,
+            tools=tools,
+            betas=["mcp-client-2025-11-20"],
+        )
+    except anthropic.BadRequestError as e:
+        # MCP server failures (rate limits, connection errors) should not
+        # block the assistant entirely — retry without MCP tools.
+        if "MCP server" in str(e) and mcp_servers is not anthropic.NOT_GIVEN:
+            logger.warning("MCP server error, retrying without MCP: %s", e)
+            response = await client.beta.messages.create(
+                model=settings.AI_MODEL_ANTHROPIC,
+                max_tokens=max_tokens,
+                system=system_blocks,
+                messages=messages,
+            )
+        else:
+            raise
 
     # Extract text from response — MCP responses can have multiple block
     # types (text, mcp_tool_use, mcp_tool_result). We join all text blocks.