MCP Is Dead. Long Live MCP
MCP was supposed to make connecting AI models to tools easier. Instead, developers started asking whether we even need it.
Over the past few months, the AI developer community has seen a growing backlash against the Model Context Protocol (MCP).
The complaints are familiar:
- Too much context consumed by tool definitions.
- Stateful architecture creates scaling problems.
- Security concerns around third-party MCP servers.
- No mature, centralized registry.
- Extra abstraction when a CLI or REST API already works perfectly well.
And that raises a legitimate question:
Is MCP actually becoming obsolete?
I don't think so.
But I do think the original vision of MCP is changing.
The most interesting part of the MCP story isn't whether developers should choose MCP, REST, or CLI.
It's what happens when a protocol moves from experimentation into production.
The backlash is real
MCP isn't being criticized because nobody uses it.
It's being criticized because people are actually trying to use it seriously.
For a developer running two local tools, MCP can sometimes feel like unnecessary ceremony. If a reliable CLI already exists, calling that CLI directly can be simpler, more transparent, and easier to debug.
For larger agent platforms, however, the requirements are very different.
You need authentication.
You need authorization.
You need observability.
You need governance.
You need scalable infrastructure.
You need predictable behavior across multiple clients and servers.
That's where the conversation becomes much more interesting.
MCP isn't necessarily dying. It's evolving.
The latest MCP specification changes the architecture significantly, particularly around stateless operation, routing, caching, authorization, and extensibility.
That matters because many of the strongest criticisms of MCP were architectural rather than conceptual.
The underlying idea — standardizing how AI systems discover and interact with tools and capabilities — is still extremely useful.
The implementation simply needs to mature.
And that's exactly what production infrastructure normally looks like.
First comes the exciting demo.
Then comes adoption.
Then come the painful production problems.
Then the architecture changes.
Eventually, nobody talks about the infrastructure anymore.
They just depend on it.
Maybe that's the real future of MCP
I don't think the future is:
MCP replaces REST.
I also don't think it's:
MCP replaces CLI.
Instead, we're probably heading toward an ecosystem where all three have a role.
A CLI can remain the best interface for humans and automation.
REST can remain the right choice for conventional application-to-application communication.
MCP can provide a standardized interface for AI agents interacting with tools and capabilities.
The interesting architecture is not about picking one winner.
It's about knowing which layer belongs where.
So, is MCP dead?
Maybe the MCP hype cycle is dying.
The protocol itself?
That's a different story.
If MCP becomes less exciting, more boring, more standardized, more secure, and easier to operate at scale, that could actually be a sign that it succeeded.
Infrastructure doesn't need to be exciting.
It needs to work.
I explored this argument in much more detail in my Medium article:
MCP Is Dead. Long Live MCP.
Read the full article on Medium:
https://medium.com/the-code-frontier/mcp-is-dead-long-live-mcp-3a9c490efdc9
I'd love to hear from other developers:
Are you actually using MCP in production, or do you still reach for REST APIs and CLIs whenever possible?
Top comments (0)