Expose MCP Server Instructions in Pydantic AI’s MCP Client

[meruzhanhovhannisyan]

Description

🧩 Feature Request: Expose MCP Server Instructions in Pydantic AI’s MCP Client

MCP servers can send server instructions along with serverInfo during initialization.
These instructions are useful for LLMs to better understand:
How to interact with the server’s tools
Known limitations or special behaviors
Other contextual guidance

Example:

{
  "capabilities": {
    "logging": {},
    "resources": { "listChanged": true },
    "tools": { "listChanged": true }
  },
  "serverInfo": {
    "name": "example.mcp.server",
    "version": "1.0.0.0"
  },
  "instructions": "Some instructions sent by server"
}

Problem
The current pydantic-ai MCP client only stores serverInfo and doesn’t expose the instructions field.
Right now, the only way to access it is by initializing the client twice, capturing the InitializeResult, and manually extracting instructions.
This is inefficient and makes it harder to incorporate valuable server context into prompts or tool logic.

Proposed Solution
Expose the instructions field in the client API, for example:
client.instructions
or include it as part of the existing server info data structure.

Why It Matters
Enables developers to easily use trusted MCP server instructions in their system prompts
Helps LLMs adapt to server-specific behaviors dynamically
Avoids redundant initialization calls

References

No response

[DouweM]

@meruzhanhovhannisyan Thanks for the report, I agree we should expose this info on MCPServer so that you can use it from a dynamic instructions function registered on the agent (that's executed after the server connection has been initialized).

A dedicated instructions property would make sense, but as MCPServer inherits from AbstractToolset, we should think about whether we'll want to introduce a new "toolset instructions that are automatically sent to the model" feature to all toolsets at some point that could end up conflicting with this feature we're adding now.

But I think that, like get_tools and call_tool, this hypothetical "toolset instructions" feature would use a get_instructions method that is passed the RunContext, and potentially also the list of ToolDefinitions. So defining instructions on MCPServer now is fine, and if we ever add that feature, we can add get_instructions as well.

Would you be up for submitting a PR? We should also include a docs example of how to use this property from a dynamic instructions function.

[mochow13]

So defining instructions on MCPServer now is fine

Hi @DouweM, if the above is still valid, it seems simple and I think I can possibly raise an MR for it—if someone else is not already working on it, of course.

[DouweM]

@mochow13 Go for it