Anthropic built MCP. Then it published the numbers showing where code beats it.
In one tested workflow, letting the agent write and run code instead of calling tools through the protocol cut context use from 150,000 tokens to 2,000. A 98.7% reduction.
Most companies would have quietly changed their own implementations and kept selling the standard. Anthropic put the limitation in public, with figures attached.
That's the part worth noticing. Not because MCP is dead (it isn't). It still earns its place for transport, authentication, discovery, and coordinating agents across vendors. But its job just got narrower, and the people who designed it said so.
The consequence lands on everyone else. "We use MCP because it's the standard" used to end the conversation in architecture reviews. Now it should start one: are we solving for interoperability, or for performance?
The reason the gap exists is simple. Models have seen vast amounts of real code and comparatively few tool-call schemas. Hand them an interface they barely know and you pay for it in tokens, latency and confused agents once the tool list grows.
If your agents are performance-sensitive and you're still defaulting to the protocol, you now need a better reason than conformity.
The creator of the standard gave you permission to question it. What's your team's reason for not taking it?
Discover more from Leverage AI for your business
Subscribe to get the latest posts sent to your email.