`ClientSessionGroup` attempts to use capabilities that were not advertised

[mr-brobot]

Initial Checks

• I confirm that I'm using the latest version of MCP Python SDK
• I confirm that I searched for my issue in https://github.com/modelcontextprotocol/python-sdk/issues before opening this issue

Description

ClientSessionGroup calls session.list_tools, session.list_prompts(), and session.list_resources() unconditionally on every connect. It never consults the ServerCapabilities returned by initialize.

For a server that doesn't implement tools, prompts, or resources, the corresponding calls return a JSON-RPC Method not found error. The MCPError is squashed and a WARNING is emitted, e.g.,

WARNING:root:Could not fetch prompts: Method not found

This seems to contradict the MCP lifecycle spec (2025-11-25, Capability Negotiation / Operation):

Both parties MUST: ... Only use capabilities that were successfully negotiated.

ClientSessionGroup should inspect session.initialize_result.capabilities and only interrogate capabilities that the server advertised, e.g.,

caps = session.initialize_result.capabilities if session.initialize_result else None

if caps and caps.prompts is not None:
    prompts = (await session.list_prompts()).prompts
    ...
if caps and caps.resources is not None:
    resources = (await session.list_resources()).resources
    ...
if caps and caps.tools is not None:
    tools = (await session.list_tools()).tools
    ...

Example Code

"""Repro: tools-only server + ClientSessionGroup emits WARNINGs."""

from __future__ import annotations

import anyio

from mcp import Client, types
from mcp.client.session_group import ClientSessionGroup
from mcp.server import Server, ServerRequestContext


async def handle_list_tools(
    ctx: ServerRequestContext, params: types.PaginatedRequestParams | None
) -> types.ListToolsResult:
    return types.ListToolsResult(
        tools=[
            types.Tool(
                name="ping",
                description="Ping",
                input_schema={"type": "object", "properties": {}},
            )
        ]
    )


async def handle_call_tool(
    ctx: ServerRequestContext, params: types.CallToolRequestParams
) -> types.CallToolResult:
    return types.CallToolResult(content=[types.TextContent(type="text", text="pong")])


async def main() -> None:
    server = Server(
        "tools-only-server",
        on_list_tools=handle_list_tools,
        on_call_tool=handle_call_tool,
    )

    async with Client(server) as client:
        # Server omits prompts/resources capabilities:
        assert client.initialize_result.capabilities.prompts is None
        assert client.initialize_result.capabilities.resources is None

        group = ClientSessionGroup()
        await group.connect_with_session(client.initialize_result.server_info, client.session)


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

Python & MCP Python SDK

Python: 3.14.4
MCP Python SDK: main @ da4c118 (also reproducible on v1.27.1)
OS: Linux 6.6.87.2-microsoft-standard-WSL2 x86_64

[scosemicolon]

This looks worth fixing at the ClientSessionGroup.connect_with_session boundary rather than only suppressing the warnings.

A clean behavior contract could be:

• if capabilities.tools is None, skip list_tools() entirely and leave the group's tool collection unchanged for that session
• do the same for prompts and resources
• only treat Method not found as recoverable for older/non-conforming servers that advertised the capability but still failed the method
• add a regression server that advertises tools only, then assert no prompt/resource requests are sent and no warning is emitted

That keeps the client aligned with negotiated capabilities while preserving a fallback path for servers with incorrect capability metadata.