Problem Description

When using the github.com/mark3labs/mcp-go library for stdio-based MCP servers, a race condition occurs after exactly 2-4 concurrent tool calls, causing:

• Error reading input: read /dev/stdin: bad file descriptor
• write |1: broken pipe errors
• Complete server failure requiring restart
• 0% success rate for subsequent tool calls

Root Cause Analysis

The MCP-Go library has a hardcoded race condition in /server/stdio.go:

// Problematic code in MCP-Go library
if baseMessage.Method == "tools/call" {
    go func() {  // ❌ RACE CONDITION: Multiple goroutines compete for same stdin
        response := s.server.HandleMessage(ctx, rawMessage)
        if response != nil {
            s.writeResponse(response, stdout)
        }
    }()
}

The Issue:

• Multiple goroutines created for concurrent tool calls
• All goroutines access the same bufio.Reader on stdin
• bufio.Reader is NOT thread-safe
• Concurrent access corrupts the file descriptor
• Results in permanent stdin corruption

After trying many workarounds for hours, Claude said:

The only reliable fix is to completely replace the MCP-Go stdio server with a custom synchronous implementation that eliminates all goroutines.

The fix worked but I would like to use the mcp-go stdio