Every app, every tool, custom code
Before MCP, connecting an AI application to an external system — a database, a ticketing tool, a filesystem, a SaaS API — meant writing integration code specific to that pairing. The code that lets Claude Desktop talk to your company's internal wiki looks nothing like the code that lets a different AI app talk to that same wiki, even though the underlying need is identical: give the model a way to read and act on wiki content. Multiply this across every AI application that might want wiki access, and every other tool each of those applications might also want to reach, and the number of integrations you need is the product of the two counts, not the sum.
That's the M×N problem: M AI applications, N tools and data sources, and in the worst case M×N distinct integrations, each maintained separately, each breaking separately when either side changes its API. It's the same shape of problem that motivated standards like ODBC for databases or POSIX for operating systems — anywhere a combinatorial explosion of custom adapters can be replaced by everyone agreeing to speak one protocol instead.