Hi, I’m developing with the Apps SDK and testing my own MCP server in Developer Mode.
What I’m trying to achieve is this flow:
User sends a natural-language question.
GPT determines that web browsing is required.
GPT performs a web search / browsing automatically.
The result of the search is passed as an input to one of my MCP tools.
The MCP tool processes it and renders a widget based on that data.
In other words: User message → Web search → Use the search result as the MCP input → Return a widget.
Before I spend time on architecture and tooling, I want to confirm whether this is currently supported by the platform.
From my testing, when Developer Mode + MCP connector is enabled, GPT seems unable to perform web browsing and instead only uses MCP tools. So I’m not sure if the intended design allows:
a conversation with web browsing,
followed by an MCP tool call using the browsing result as input,
all within the same chat session.
Is this possible today?
Or does the browsing capability need to be exposed separately (e.g., as another MCP tool instead of using built-in web search)?
Any clarification or recommended architecture would be appreciated!
Currently, when Developer Mode is enabled, the model cannot combine built-in browsing with MCP tools in the same flow. When MCP is activated, the model usually gives those tools priority and doesn’t use the integrated web search. The safer and more reliable architecture is to expose a custom browsing tool through MCP if your workflow relies on “browse → feed into tool.”
Thanks a lot for the explanation — that matches what I was observing in Developer Mode.
I have a follow-up question to make sure I fully understand the limitations:
Is this restriction only for Developer Mode, or does it also apply in a “normal” / production ChatGPT environment?
Concretely, in a regular chat where:
the user has enabled my MCP connector (so my tools are available), and
built-in browsing is also available for that model,
is the model able to use both in the same conversation flow? For example:
User question → model uses built-in browsing → then calls one of my MCP tools using the browsing result as input → then replies to the user
Or does enabling an MCP connector in a regular chat session also cause the model to prefer MCP tools and effectively stop using built-in web search, similar to what happens in Developer Mode?
Any clarification on how this behaves outside Developer Mode would be super helpful. Thanks again!
User message → Web search → Use the search result as the MCP input → Return a widget.