diff --git a/src/llama_stack/providers/inline/batches/reference/batches.py b/src/llama_stack/providers/inline/batches/reference/batches.py
index 79dc9c84c..b6f201fae 100644
--- a/src/llama_stack/providers/inline/batches/reference/batches.py
+++ b/src/llama_stack/providers/inline/batches/reference/batches.py
@@ -419,8 +419,8 @@ class ReferenceBatchesImpl(Batches):
                         )
                         valid = False
 
-                    if (body := request.get("body")) and isinstance(body, dict):
-                        if body.get("stream", False):
+                    if (request_body := request.get("body")) and isinstance(request_body, dict):
+                        if request_body.get("stream", False):
                             errors.append(
                                 BatchError(
                                     code="streaming_unsupported",
diff --git a/src/llama_stack/providers/utils/tools/mcp.py b/src/llama_stack/providers/utils/tools/mcp.py
index 48f07cb19..a271cb959 100644
--- a/src/llama_stack/providers/utils/tools/mcp.py
+++ b/src/llama_stack/providers/utils/tools/mcp.py
@@ -15,7 +15,7 @@ from mcp import types as mcp_types
 from mcp.client.sse import sse_client
 from mcp.client.streamable_http import streamablehttp_client
 
-from llama_stack.apis.common.content_types import ImageContentItem, InterleavedContentItem, TextContentItem
+from llama_stack.apis.common.content_types import ImageContentItem, InterleavedContentItem, TextContentItem, _URLOrData
 from llama_stack.apis.tools import (
     ListToolDefsResponse,
     ToolDef,
@@ -49,7 +49,9 @@ async def client_wrapper(endpoint: str, headers: dict[str, str]) -> AsyncGenerat
         try:
             client = streamablehttp_client
             if strategy == MCPProtol.SSE:
-                client = sse_client
+                # sse_client and streamablehttp_client have different signatures, but both
+                # are called the same way here, so we cast to Any to avoid type errors
+                client = cast(Any, sse_client)
             async with client(endpoint, headers=headers) as client_streams:
                 async with ClientSession(read_stream=client_streams[0], write_stream=client_streams[1]) as session:
                     await session.initialize()
@@ -137,7 +139,7 @@ async def invoke_mcp_tool(
             if isinstance(item, mcp_types.TextContent):
                 content.append(TextContentItem(text=item.text))
             elif isinstance(item, mcp_types.ImageContent):
-                content.append(ImageContentItem(image=item.data))
+                content.append(ImageContentItem(image=_URLOrData(data=item.data)))
             elif isinstance(item, mcp_types.EmbeddedResource):
                 logger.warning(f"EmbeddedResource is not supported: {item}")
             else:
diff --git a/src/llama_stack/testing/api_recorder.py b/src/llama_stack/testing/api_recorder.py
index 84407223c..e0c80d63c 100644
--- a/src/llama_stack/testing/api_recorder.py
+++ b/src/llama_stack/testing/api_recorder.py
@@ -40,7 +40,9 @@ from openai.types.completion_choice import CompletionChoice
 from llama_stack.core.testing_context import get_test_context, is_debug_mode
 
 # update the "finish_reason" field, since its type definition is wrong (no None is accepted)
-CompletionChoice.model_fields["finish_reason"].annotation = Literal["stop", "length", "content_filter"] | None
+CompletionChoice.model_fields["finish_reason"].annotation = cast(
+    type[Any] | None, Literal["stop", "length", "content_filter"] | None
+)
 CompletionChoice.model_rebuild()
 
 REPO_ROOT = Path(__file__).parent.parent.parent.parent