Q

Will MCP survive 18 months?

Probably yes. Here's why — and what could kill it.

Will MCP survive 18 months?

Short answer: Probably yes. The momentum is real, the spec is clean, and the cross-vendor support is unprecedented for an Anthropic-led standard. Microsoft and OpenAI have both adopted it.

But survival isn’t guaranteed. Here’s the honest read.

Why MCP is likely to stick

1. Cross-vendor adoption. Microsoft Copilot Studio, OpenAI’s Apps SDK (which builds on MCP-style servers), and Anthropic Claude all support MCP today. That’s unusual cross-vendor agreement.

2. Openly specified, but still Anthropic-led. Anyone can implement MCP, which helps adoption. Governance is not yet under a neutral standards body, so that remains a risk (more on that below).

3. Real momentum. Hundreds of MCP servers in registries, with new ones added weekly. Network effects are kicking in.

4. Practical design. The protocol solves real problems (stateful context, tool discovery, structured I/O). It’s not academic.

5. No clear competitor at the same scale. A2A is younger and narrower. Apps SDK builds on top of MCP rather than replacing it. Open Agent Protocol has minimal traction.

What could kill it

Honest list of survival risks:

Risk 1: Microsoft or OpenAI fork it. If a major vendor decides MCP doesn’t fit their roadmap and they have the market position to push their own, they could ship “MCP++” or “MCP for X”. This has killed open standards before.

Risk 2: A simpler protocol emerges. MCP is currently auth-heavy and stateful. If a leaner alternative ships and gets adoption, MCP becomes the JavaScript-frameworks-of-2014 — too many features, ready for replacement.

Risk 3: Breaking changes in future spec revisions. The MCP spec is versioned by date and still moves quickly. If a future revision introduces aggressive breaking changes, fragmentation is a real risk. We’ve seen this kill open standards before.

Risk 4: Anthropic loses momentum. If Anthropic the company stumbles (acquisition, regulatory issues, etc.), MCP doesn’t have a neutral standards body yet. It would need to find one fast.

Risk 5: Major security incident. A high-profile MCP-mediated breach could trigger overreaction and ecosystem panic.

My probability estimate

If you forced me to put a number on it:

Outcome18-month probability
MCP is still the dominant agent-tool protocol~70%
MCP coexists with 1-2 alternatives without dominating~20%
MCP fragments / replaced by an alternative~10%

Whatever happens, the ideas in MCP (open tool-server architecture, host-server separation, structured context) will persist. The specific spec might evolve.

What to do regardless

Don’t bet your entire stack on the spec being permanent. Bet on the patterns being permanent.

That means:

  • Build MCP servers because the work is portable even if the protocol changes
  • Use MCP-aware hosts (Claude Desktop, Cursor, etc.) because they’ll likely adopt successors
  • Avoid deep MCP-specific UI/UX coupling that would be hard to migrate
  • Watch the v2 RFC closely — that’s where the next 18 months get decided

My honest editorial take

My take: MCP is the cleanest widely-adopted tool-connection protocol I’ve seen in this wave of LLM apps. I would build on it today without hesitation. If something better comes along in 18 months, the migration cost will be modest because the patterns transfer.

The bigger risk than “MCP dies” is “you spent 18 months waiting for the protocol to settle and your competitors shipped”.