Agent run calling MCP list tools every time before calling the model provider

[psg2]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

When running an agent connected with a MCPServerStreamableHTTP I've noticed in our Datadog traces that the agent calls MCP list tools every time prior to sending a request to the model provider even a single run.

Also if the list tools fails the entire run fails with an McpError or HTTPStatusError.

Earlier today our agent failed with:

Server error '520 <none>' for url 'https://api.fireflies.ai/mcp'
For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/520

In the example code we can see two log calls like:

DEBUG:mcp.client.streamable_http:Sending client message: root=JSONRPCRequest(method='tools/list', params=None, jsonrpc='2.0', id=1)
DEBUG:mcp.client.streamable_http:Sending client message: root=JSONRPCRequest(method='tools/list', params=None, jsonrpc='2.0', id=2)

Slack thread: https://pydanticlogfire.slack.com/archives/C083V7PMHHA/p1761072691821499

Example Code

# type: ignore
import asyncio
from datetime import date

from pydantic_ai import (
    Agent,
    RunContext,
)
import dotenv
import os
from pydantic_ai.mcp import MCPServerStreamableHTTP
import logging

logging.basicConfig(level=logging.WARNING)
mcp_logger = logging.getLogger("mcp.client.streamable_http")
mcp_logger.setLevel(logging.DEBUG)

dotenv.load_dotenv(".env.local")

FIREFLIES_API_KEY = os.getenv("FIREFLIES_API_KEY")

fireflies_server = MCPServerStreamableHTTP(
    "https://api.fireflies.ai/mcp",
    tool_prefix="mcp_fireflies",
    headers={"Authorization": f"Bearer {FIREFLIES_API_KEY}"},
    timeout=30,
)

weather_agent = Agent(
    "bedrock:us.anthropic.claude-3-5-haiku-20241022-v1:0",
    system_prompt="Providing a weather forecast at the locations the user provides.",
    toolsets=[fireflies_server],
)


@weather_agent.tool
async def weather_forecast(
    ctx: RunContext,
    location: str,
    forecast_date: date,
) -> str:
    return f"The forecast in {location} on {forecast_date} is 24°C and sunny."


output_messages: list[str] = []


async def main():
    user_prompt = "What will the weather be like in Paris on Tuesday?"

    result = await weather_agent.run(user_prompt=user_prompt, deps=None)

    print(result.output)


if __name__ == "__main__":
    asyncio.run(main())

Python, Pydantic AI & LLM client version

Python 3.12.11
Pydantic AI Slim 1.2.1
AWS Bedrock

[DouweM]

I ended up having to revert the change: #3228. We'll have to properly handle tool change notifications: #3227 (comment).