feat: add fault tolerance for MCP server connections

[Edison-A-N]

Overview

Adds fault tolerance support for handling unavailable MCP servers when using shared MCP config files.

Motivation

When sharing MCP config files across teams/projects, some MCP servers may become unavailable due to poor maintenance or temporary issues. Currently, a single failed server connection causes the entire adapter to fail, blocking access to all other working servers.

Discussion point: Should MCP server availability be managed solely by the config provider, or should the SDK help manage failures gracefully?

Changes

• fail_fast parameter (default: True):

  • True: Maintains backward compatibility - any connection failure raises an exception
  • False: Skips failed connections and continues with available servers

• on_connection_error callback: Optional error handler called for each failed connection with (server_params, exception) as arguments

• failed_connections tracking: List of tuples (server_params, exception) for transparency

• Individual connection handling: Changed from batch connection to individual try-catch for better fault isolation

Usage Example

def log_connection_error(params, exception):
    logger.warning(f"Failed to connect to {params}: {exception}")

with MCPAdapt(
    [server1_params, server2_params, server3_params],
    adapter=MyAdapter(),
    fail_fast=False,  # Continue even if some servers fail
    on_connection_error=log_connection_error
) as tools:
    # Only tools from successfully connected servers are available
    # Check adapter.failed_connections for error details
    pass

Testing

Comprehensive test coverage added in test_core_fault_tolerance.py covering:

• Default fail_fast=True behavior
• Fault tolerance with fail_fast=False
• Error callback functionality
• Failed connection tracking
• Both sync and async context managers