MCP Prompts API support

[sergii-vashchyshchuk]

MCP defines a nice Prompts concept:
https://modelcontextprotocol.io/docs/concepts/prompts

But openai-agents seems don't support it yet. Any plans implementing it natively?

[rm-openai]

We could add it to the MCPServer class so you can fetch prompts. Would anything else be useful?

[sergii-vashchyshchuk]

I thought the whole idea is Agent/Runner to support it.

I.e. the user sends some simple, generic prompt to an agent ("Analyze $TSLA"), the agent gets prompts from a MCP server(s), tries to match with one of templates (using LLM or BERT), then if match found requests a refined detailed prompt from the MCP server ("Find the latest news and financial reports for $TSLA stock, analyze blahbla and provide blahblah report including blhalblah"), then the agent sends the refined prompt to an LLM with tools from MCP server.

I know, it is possible to workaround by creating a specific "Analyse stock ticker" agent with proper system instructions and MCP server, but having dozens of single purposes specialized agents won't scale well.

[rm-openai]

I don't think we'd support this use case in the SDK. Choosing a prompt is generally a carefully thought out process, and the LLM/Bert approach isn't a universally useful one. In addition, the MCP spec actually frames prompts as a user-visible choice (i.e. the user picks the prompt via a UI).

[sergii-vashchyshchuk]

Yeah, I reread the spec and realized that they suggest using prompts for UI user suggestions, which significantly limits the potential of the prompts API.

LLMs are often quite naive when suggesting tools based solely on user prompts. To get the best results, you need to guide the LLM using different techniques to determine what needs to be done to answer the user query and what tools might be required. Having a feature like the prompts API is a powerful mechanism, as it allows agents/LLMs to refine the user query in a more deterministic and precise way—aligned with how the developers of an MCP server designed it. MCP server developers also have a better understanding of the tools they expose and the best ways to utilize them.

I agree that modifying or replacing the user query should be done carefully. We just need to ensure that a provided prompt template is used only if it fits the user query; otherwise, the original query should be used. This way, we get the best of both worlds—leveraging a prompt template when it’s a good match and sticking with the user’s query when it isn’t.

Actually, I built a proof of concept for this approach in my agents framework, and it works pretty well.

[github-actions]

This issue is stale because it has been open for 7 days with no activity.

[github-actions]

This issue was closed because it has been inactive for 3 days since being marked as stale.