Summary

Fixes #2527 — MCPServer.__init__ called configure_logging() which used logging.basicConfig() to install a RichHandler on the root logger at INFO level. This caused every third-party library (httpx, urllib3, openai, qdrant-client, …) to emit INFO records through Rich → stderr. In stdio MCP servers spawned by Node.js hosts (Claude Desktop, Cline, qwen-code), the kernel's stderr SNDBUF (8 KB on macOS) fills under modest log volume and write(2) blocks the asyncio event loop, deadlocking the server.

Changes

• configure_logging() now configures logging.getLogger("mcp") instead of the root logger via basicConfig.
• Sets mcp_logger.propagate = False so records don't bubble up to any root handlers.
• Made idempotent: skips setup if the mcp logger already has handlers, allowing applications to configure logging before constructing MCPServer.
• Default log level changed "INFO" → "WARNING" — reduces noise from transitive dependencies in production/stdio deployments. Users can pass log_level="INFO" (or "DEBUG") explicitly when they want verbose output.
• MCPServer.__init__ log_level parameter default also changed to "WARNING".
• Added regression tests in tests/issues/test_2527_fastmcp_logger_pollution.py.

Before / After

Before (broken):

def configure_logging(level="INFO"):
    handlers = [RichHandler(console=Console(stderr=True), ...)]
    logging.basicConfig(level=level, handlers=handlers)  # ← root logger!

After (fixed):

def configure_logging(level="WARNING"):
    mcp_logger = logging.getLogger("mcp")
    if mcp_logger.handlers:
        return  # idempotent
    mcp_logger.addHandler(RichHandler(console=Console(stderr=True), ...))
    mcp_logger.setLevel(level)
    mcp_logger.propagate = False  # don't reach root

Test plan

• uv run pytest tests/issues/test_2527_fastmcp_logger_pollution.py -v — 5 new regression tests, all pass
• uv run pytest tests/server/mcpserver/test_server.py -v — 89 existing tests, all pass

🤖 Generated with Claude Code