Everyone is reading the MCP spec. Almost nobody is reading the distribution map.
That is the mistake. The technical specification is the least interesting thing about the Model Context Protocol. It is JSON-RPC with some primitives bolted on. Useful, clean, boring. The interesting question is not how the wire format works. It is who owns the tool surface your agent reaches for by default. That is a distribution fight dressed up as a standards conversation, and most engineering leaders are still reading the wrong document.
Start with the number that reframes everything. The SDK was downloaded roughly 100,000 times that first month. By March 2026, that number had reached 97 million monthly downloads, a 970x increase in 18 months. For comparison, the React npm package took approximately 3 years to reach 100 million monthly downloads. MCP achieved comparable scale in 16 months.
No protocol wins that fast on the strength of its schema. It wins on defaults.
Adoption is a distribution event, not a technical one
Here is the tell. The moment MCP stopped being one option among many was not a spec release. It was a platform commitment. Infrastructure standards only become infrastructure when all major players adopt them. Prior to that point, they are just one of several competing approaches. MCP crossed that threshold in 2025 when OpenAI committed to MCP support, breaking the provider-specific tool format fragmentation.
Read that carefully. The fragmentation that ended was not a technical fragmentation. Every vendor already had function calling that worked fine. What ended was the strategic standoff over whose format would become the shelf that everyone stocks. Once OpenAI, Google, Microsoft, and the rest lined up, the format debate was over and the surface debate began.
And the surface debate is the whole game. OpenAI, Google, Microsoft, and Salesforce all shipped support within 13 months. Most marketing teams haven’t noticed yet, but the vendors building MCP servers have clear incentive to get there. That incentive is not interoperability for its own sake. It is placement. If your product is the default MCP server an agent hits when a user says “check my calendar” or “pull the invoice,” you own that intent. The competitor who ships a technically superior server two quarters later does not get a second look, because the agent already reached for you.
This is the same dynamic that made the App Store, npm, and the browser extension marketplace matter far more than the underlying APIs ever did. The API is a commodity. The default placement is the moat.
The registry is where the money is
Watch where the roadmap energy is going. It is not going into the transport layer. It is going into the catalog. Anthropic’s MCP roadmap for 2026 focuses on three areas: enterprise authentication, multi-agent coordination, and the MCP registry, a curated, verified server directory with security ratings. Each of these addresses observed gaps in the current ecosystem as enterprise adoption scales.
A curated, verified, rated directory is not a protocol feature. It is a distribution channel with an editorial policy. Whoever runs that registry decides which servers surface first, which get the trust badge, and which get buried. That is retail merchandising for agents. The open question the market has not answered yet is how official and downstream registries handle trust, curation, and stale packages. Whoever answers it owns the shelf.
The enterprise version of this is already visible. Teams are not asking for a better spec. Enterprise teams will demand registry, approval, and audit layers before broad write access. Translation: the buyer wants to control which servers become the default in their own four walls. That is a distribution and governance decision, not a standards decision.
The Cost of Doing Nothing is losing the default
Here is the CODN math for anyone shipping agents in production. The cost of doing nothing on MCP is not a missing integration you can add later. It is watching a competitor become the default server for your category while you are still debating whether the protocol is mature enough. Defaults compound. Once an agent is trained, prompted, and wired to reach for someone else’s server, you are not fighting a feature gap. You are fighting muscle memory at the model layer.
So stop auditing the spec and start auditing the surface. Ask three questions. When your agent needs a tool in your category, whose server does it reach for first. Who controls the registry your enterprise buyers trust. And what does it cost you, per quarter, to not be the default.
The protocol is settled. The distribution is not. That is where the next two years get decided, and it will not be decided by the people reading the spec.