Why Developers Are Abandoning MCP: The Costly Reality

The MCP (Model-Command Protocol) is under fire, facing harsh criticism from developers who are questioning its viability. The conversation, reignited by Eric Holmes' viral article, "MCP is dead, long live the CLI," has spread through forums and developer circles. It's a stark reminder: enterprises don't buy AI. They buy outcomes.

The Hidden Costs of MCP

MCP isn't just a protocol. It's a financial model in disguise, dressed as JSON-RPC. Developers are now realizing that each interaction with MCP incurs a 'manifest tax' for every tool used. This isn't a mere technicality, it's a real cost that adds up fast. Pieter Levels' comparison to the controversial LLMs.txt highlights the skepticism around MCP's practical value.

The tool-count cliff is another stumbling block. Large tool lists are causing a collapse in selection efficiency due to attention saturation. It forces enterprises to question whether the ROI justifies the complexity. The consulting deck says transformation. The P&L says different.

Security and Operational Failures

The security risks associated with MCP can't be ignored. Ox Security's warnings about STDIO execution pushing sanitization responsibility onto developers are serious. The 'expected behavior' model introduces risks like command injection unless sanitized pre-launch. This isn't just a technical flaw, it's a potential breach waiting to happen.

Operationally, stdout corruption remains a significant concern. Stray logs can break JSON-RPC framing, creating another layer of complexity. The practical guidance is clear: limit tools, sanitize at every controllable layer, and ensure logging is directed to stderr. Yet, the gap between pilot and production is where most fail.

The Uncertain Future of MCP

Will progressive disclosure become the norm? How will Anthropic, a key player, adapt its stance? These are the unanswered questions hanging over MCP's future. Enterprises need clarity on whether the tool-selection cliffs are intrinsic or simply a result of flawed training models.

In practice, MCP's role in multi-user or enterprise governance could justify its overhead. But the path forward demands a clear understanding of its total cost of ownership. Can MCP evolve to meet these challenges, or will it remain a relic of what could have been?

The ROI case requires specifics, not slogans. For MCP to remain relevant, it needs to prove its worth beyond theoretical promise. The onus is on developers and enterprises to decide if MCP's operational complexities and hidden costs are a risk worth taking.