MCP Server instructions are currently ignored

[bocytko]

Please read this first

• Have you read the docs?Agents SDK docs
• Have you searched for related issues? Others may have had similar requests

Question

MCP Servers define instructions, for example the aws-documentation-mcp-server.

Instructions are currently ignored in the Agent execution flow. It only reads mcp tools and converts these to LLM tool descriptions. Is this an explicit design decision?

[rm-openai]

Not explicit, I didn't realize this existed. From the docs, looks like instructions is optional and it says

For example, this information MAY be added to the system prompt.

We could add a way for you to fetch the instructions, and then you can choose to append that to the system prompt. Open to other ideas too.

[bocytko]

We could add a way for you to fetch the instructions, and then you can choose to append that to the system prompt.

This sounds intuitive enough to me. Shall I file a separate feature request?

[rm-openai]

Nope, we can use this issue. PR also welcome!

[DanieleMorotti]

Hi @bocytko, if you haven’t started on this yet I’d be happy to implement this feature.
I found that the instructions are included in the instructions field of the session‐initialization response. We could store it on the server and inject them into the prompt where needed.

[bocytko]

@DanieleMorotti please go ahead. Thank you in advance!

[DanieleMorotti]

@rm-openai I can do something like this:

server_result = await session.initialize()
self.mcp_instructions = server_result.instructions or None

in src/agents/mcp/server.py at line 125. But how / where to pass this variable only for the tools related to a specific server? In the FunctionTool object?

[rm-openai]

@DanieleMorotti wdym? I don't think you need to do anything else. Once you do that, users will be able to do this:

async with MCPServerStreamableHttp(...) as server:
  agent = Agent(..., mcp_servers=[server], instructions=f"...MCP instructions: {server.instructions}")

[DanieleMorotti]

Ok sorry, I thought it should be injected to the prompt somehow but instead users will directly choose how to adopt instructions.

[rm-openai]

Ah yup, prefer to leave it to users - less magic that way.