When my MCP tool returns image content blocks (type: “image”), ChatGPT connector invokes the tool successfully but the tool response body appears as {} (empty).
The exact same tool and server return valid image output in MCP Inspector.
Actual behavior
-
Tool is discovered and invoked successfully.
-
ChatGPT connector UI/log shows a successful tool call.
-
Response body is {} for image-block tool output.
-
No image context is available to the model for analysis.
Reproduction
-
Start MCP SSE server.
-
Connect ChatGPT connector to the server.
-
Call tool take_screenshot_image (returns MCP image block).
-
Observe empty {} response body in ChatGPT connector.
-
Test same tool in MCP Inspector: image is returned correctly.
Tool outputs
Tool A: take_screenshot
Returns JSON payload:
-
imageBase64
-
mimeType
-
width
-
height
-
byteLength
This is visible in ChatGPT connector (but very large and not practically interpreted as image context).
Tool B: take_screenshot_image
Returns MCP image block:
In ChatGPT connector this appears as {}.Why this matters
This blocks core MCP multimodal workflows. Screenshot/image tools are common, and connector users currently cannot rely on image tool outputs for visual analysis.
Request
-
Bug fix: support MCP image content blocks in ChatGPT connector tool responses.
-
If not yet supported, please document this limitation explicitly in connector/MCP docs.
-
Feature request: first-class multimodal connector path for MCP tool image outputs.
Additional context
Similar client-side issue/fix pattern exists in Cline:They fixed it
I can share sanitized logs and a minimal reproducible MCP server if needed.
-