Playground Bug - MCP Server URL Gets Redacted When Saving Prompt in Playground

API
,
1

I’m encountering a consistent issue when using the Playground with a custom MCP server integration.

  • MCP server is running at a valid public HTTPS URL, e.g.,
    https://my-server.azurewebsites.net/sse/
  • Tools are: search and fetch
  • Auth token is correctly provided
  • Server works perfectly when used directly in Prompt Playground before saving

When I save the prompt in the Playground :

  • The server URL is automatically redacted in the interface
from: https://my-server.azurewebsites.net/sse/
to:   https://my-server.azurewebsites.net/<redacted>

Unable to load tools  
Check the server URL and verify your authentication details.

The image shows two side-by-side screenshots of an API tool configuration interface, with the left side successfully displaying available tools after entering an API key, while the right side shows an authentication error and a message stating it is unable to load tools. (Captioned by AI)
The image shows two side-by-side screenshots of an API tool configuration interface, with the left side successfully displaying available tools after entering an API key, while the right side shows an authentication error and a message stating it is unable to load tools. (Captioned by AI)1253×667 42.5 KB

8 Likes
5

I ran into the same issue. I created a prompt with an MCP server. It works initially but once I close and re-open the prompt, the path changes to / and the prompt is unable to connect to the MCP server.

2 Likes
6

The playground sends the request to make a prompt with POST and the URL you specify. The return from that call literally places <redacted> as the path when sending back the “server_url” field of the JSON to its POST API.

Then it’s prompt listing and individual prompt ID API request also returns the same path.

It will work in that same playground instance, but merely reusing the playground URL https://platform.openai.com/chat/edit?prompt=pmpt_... you get the bad path in the function.

It seems the underlying intention is to never present that URL data again, like an API key. That is, if the prompt is working in a call by using it’s ID

Guess you could run your MCP “prompt” if you’re into that kind of thing.

NOPROMPT

Otherwise, “get code” never having created a prompt.

  • See how to make the MCP tool call.
  • See how to not use a server state, because even playground programmers rightfully don’t use previous_response_id.
  • See another bug where the system message is neither in “input” or in “instructions” of code if you haven’t edited that field again.

Then figure this out: update the model to one that wasn’t cancelled, save a new prompt version. The API call is made with the prompt ID but the system message box is toast.

The image shows a user interface for an AI prompt creation tool, with a notification about the unavailability of the gpt-4.5-preview model and an exchange about AI tool-calling capabilities. (Captioned by AI)
The image shows a user interface for an AI prompt creation tool, with a notification about the unavailability of the gpt-4.5-preview model and an exchange about AI tool-calling capabilities. (Captioned by AI)1196×600 29.3 KB

7

same issue here. Usage is not clear

8

Hello. have you solved this problem?

9

I’ve shared this thread with my account rep at OpenAI. They say the Engineering team is aware of the issue. I’ll report back if/when I get an update and/or a workaround.

1 Like
10

I’ve been in contact with OpenAI reps and support. There is unfortunately no workaround, but they’re working on a fix, hoping to have it sometime this week or next.

3 Likes
11

Just jumping in on this thread as I am experiencing the same issue currently when using an MCP server cloudflare connection on a audio project, the moment I update the project it is unable to load tools.

13

Hi! I experienced the same issue here. I’d like to add my log from my Cloudflare MCP. Hope it helps!

Screenshot 2025-08-28 at 16.05.10
Screenshot 2025-08-28 at 16.05.102354×938 260 KB

15

Alright, OpenAI hasn’t made progress on this, so I’m going to share my super hacky workaround.

First, we’ll just have to have our MCP server listen at /<redacted>. Hopefully you have an implementation that allows you to do that. Second, for authentication, we’ll have to rely on the name of our Custom Header(s) for security, since their values will automatically be set to REDACTED. Maybe set a couple of them, or set the name to something hard to guess. This means you’ll need to be able to implement some custom authentication where you can check for the existence of these header(s) as your security measure.

If you can do these things, you should be able to workaround this issue.

16

Alternative: find “Prompts” useless and just send the API parameters in the call.

17

Misery loves company, glad I’m not the only one scratching my head at this!

I like @Tom_Duffield’s approach! My solution was to use a subdomain (e.g. mcp.[domainname].com), because only pathnames seem to get the redacted treatment.

This makes me wonder how many people are using the custom MCP feature in Responses API.

2 Likes