Is MCP Actually Dead, or Do We Just Like Complaining?

Open any developer timeline this month and you'll find someone burying the Model Context Protocol. "MCP is dead." "MCP was a mistake." "Nobody uses it right." Scroll a little further and you'll find the same people shipping MCP servers. That contradiction is the whole story, so let's sit with it.

MCP is, right now, both the most widely adopted standard for connecting AI agents to tools and the one developers most love to dunk on. When something is simultaneously everywhere and universally complained about, "is it dead?" is the wrong question. The right one is: what are people actually mad about, and does it matter?

The complaints are real

I want to be fair to the critics, because they're not wrong about the specifics:

• Context bloat. Some MCP servers dump 40k+ tokens of tool descriptions into the model's context before you've done anything. You're paying — in money and attention budget — to tell the model about tools it may never call.
• Security was an afterthought. This is the sharp one. There's even an arXiv taxonomy paper cataloguing real MCP faults. The protocol recommends authentication but doesn't enforce it — an opt-in security posture in a world that needs secure-by-default.
• Distribution is immature. npm has a registry. PyPI has a registry. MCP has... GitHub stars and README files. For something pitched as the backbone of agent tooling, "find a server by vibes" doesn't scale.

None of those are nitpicks. If you've wired up a few servers, you've felt at least two of them.

But "annoying" isn't "dead"

Here's where I get off the funeral train.

A standard doesn't die because it's irritating. It dies when something better replaces it — and right now, nothing has.

The alternatives floating around solve one slice and lose the rest: session management, dynamic tool discovery, bidirectional communication. MCP's real value isn't that it's elegant — it's that it's the only open, vendor-neutral option in the room. Kill it and you don't get a cleaner protocol; you get a vacuum, and vacuums in this industry get filled by whatever the biggest vendor ships. I'd rather debug a flawed open standard than be locked into a perfect proprietary one.

My actual take

MCP is in its awkward adolescence. The pattern is old and familiar: a thing gets adopted faster than it gets good, everyone feels the rough edges at once, and the discourse swings to "it's over" right as it's quietly becoming infrastructure. JavaScript went through it. REST went through it. Kubernetes is arguably still in it.

What I'd actually want isn't a funeral — it's a 2.0 that treats the complaints as a spec: enforce auth, lazy-load tool descriptions so context isn't pre-bloated, and build a real registry. The "MCP is dead; long live MCP" framing some people are using is closer to right than either extreme.

So no, I don't think it's dead. I think it's load-bearing and unfinished — which is the least glamorous, most important phase a technology can be in. Complaining about it is the sign it matters. Nobody writes angry threads about tools they've already abandoned.