Image response from an MCP server

[stneng]

#1898 adds support for image outputs in functions; the existing fallback logic in MCP tools below is outdated.

openai-agents-python/src/agents/mcp/util.py

Lines 217 to 227 in ba55bbd

	# Fall back to regular text content processing
	# The MCP tool result is a list of content items, whereas OpenAI tool
	# outputs are a single string. We'll try to convert.
	if len(result.content) == 1:
	tool_output = result.content[0].model_dump_json()
	elif len(result.content) > 1:
	tool_results = [item.model_dump(mode="json") for item in result.content]
	tool_output = json.dumps(tool_results)
	else:
	# Empty content is a valid result (e.g., "no results found")
	tool_output = "[]"

Resolves #2148 and https://community.openai.com/t/image-response-from-an-mcp-server-with-agents-sdk/1269441

[seratch]

Thanks for sending this. Indeed, this could work well now. I will do thorough testing early next year.

[seratch]

I ran a few local codex reviews and this feedback seems to be valid:

• [P1] Tracing output double-encodes MCP tool results — src/agents/mcp/util.py:250-252
  When recording MCP tool invocations, current_span.span_data.output is now set via json.dumps(tool_output) even though tool_output is already the final return value (string for structured
  content, or dict/list for text/image content). This extra dump double-encodes structured content (producing strings like "{"foo":1}") and converts dict/list outputs to JSON strings with
  different quoting than the returned value, causing trace snapshots to be incorrect for any MCP tool call. Span data should store the raw tool output instead of re-encoding it.

[stneng]

Thanks for the feedback, just fixed it.