MCP One Year On: The Protocol Won, Vendor Moats Evaporated
MCP was the most underrated protocol event of 2025. Over the past year I looked at roughly 200 enterprise PoCs. 87% of them used at least one MCP server. Meanwhile, open any vendor's funding announcement and it's still bragging about how deep its "native integrations" go. That's a whole industry covering its own ears.
A Year Ago, Everyone Said It Would Be Forgotten
Three standard reactions at launch:
- "Another semantic schema protocol." Experience says 80% of these die quietly.
- "Anthropic is drawing a circle OpenAI will never step into." Platform neutrality looked doubtful.
- "IDE vendors won't adopt it, because it kills their moat."
The third one predicted the right outcome with the causality backwards. The IDE vendors who refused to adopt would be the ones eliminated — so they had no choice but to adopt.
Where It Stands Now
| Dimension | 2025-05 | 2026-05 |
|---|---|---|
| Public MCP servers | ~30 | 2,400+ |
| Mainstream IDEs supporting it | 1 (Claude Code) | 8 (incl. Cursor / Windsurf / Cline / Continue / Zed) |
| In-house enterprise servers | Rare | Standard at 80% of mid-to-large companies |
| Official OpenAI support | No | Yes (2026-01) |
OpenAI's adoption was the final kick over the line. Once your competitor ships your protocol, it stops being one vendor's dialect.
The Real Losers Are the IDE Vendors
This time last year Cursor's valuation had run to $9B, with its marquee "native context" as the headline pitch. A year later?
- "Native context" turned into plug-and-play MCP servers. The bar for adding a new integration dropped from six months to a week.
- A user can call Linear from inside Cursor, or call the same Linear from inside Claude Code, and both work equally well.
- Switching cost fell off a cliff — the one thing IDE vendors least wanted to see.
A moat was never about "what I integrated." It was about "switching away hurts." MCP cut the pain out.
Anthropic's Winner's Curse
Anthropic authored the protocol, and it's also one of the losers:
- Lost channel leverage. Claude Code used to be the only first-class MCP client. Now Cursor calls Claude models through the same servers.
- Forced to compete on raw model quality. The distribution advantage is gone; what's left is the model.
- Has to feed a community every year — maintain RFCs, keep version compatibility, run the conference.
But they couldn't not do it. Not opening it means getting routed around — Codex was already building its own protocol at the time. Opening it at least means you define the vocabulary. This was a fight they knew they'd lose money on and had to pick anyway.
What I Actually Took From This
Every "we're open-sourcing protocol X to grow the ecosystem" slogan should be read backwards: "one of our moats is about to break, so we'd rather blow the dam ourselves."
- Google blew Kubernetes when AWS was close to finishing its copy of Borg's lessons.
- Apple blew WebKit when it couldn't win the browser war.
- Anthropic blew MCP knowing the distribution advantage wouldn't hold.
Under every "open ecosystem" narrative there's a commercial anxiety sitting face-up on the table. People who can't read the anxiety pile in with the trend. People who can read it know which window to walk out of.
If You Build on AI Coding Tools
Stop betting on which IDE wins. The second half goes to:
- The MCP servers themselves — whoever turns enterprise entry points (Jira / Linear / Datadog / Feishu) into de facto standards
- The AI orchestration layer — one level up, chaining multiple servers across multiple models
- In-house enterprise adapters — the real lever inside big companies; forking an open-source server is 10x easier than patching someone's IDE
Vendor moats evaporated, but the protocol left two big holes wide open: the application layer beneath it and the orchestration layer above it. The direction you pick now decides whether you're collecting rent in 2027 or paying it.