Six months ago, if you asked a roomful of engineers what Model Context Protocol was, half would have said “the thing Anthropic shipped” and the other half would have stared at you. Today, MCP is running in production at thousands of companies, the Linux Foundation owns the spec, and the install count just crossed 97 million. That number is from March 2026. By the time you finish reading this, it is higher.
Here is what changed, why it matters, and what your team should actually do about it.
The short version of MCP
- MCP is an open protocol for connecting AI models to tools, data sources, and external systems.
- Anthropic introduced it in late 2024. It went open under the Linux Foundation in December 2025 as part of the new Agentic AI Foundation.
- Think of it like USB for AI. One spec, many implementations, no custom adapter for every integration.
- Every major model vendor now supports it. Claude, GPT-5.4, Gemini 3.1, and most open-weight models can speak MCP out of the box.
The numbers driving the story
- 97 million installs as of March 2026, up from roughly 12 million a year earlier.
- Forrester projects 30% of enterprise application vendors will ship their own MCP server in 2026.
- Gartner projects 40% of enterprise applications will embed task-specific AI agents by year end, up from less than 5% in 2025. That is the steepest adoption curve they have measured for any enterprise tech category in the past decade.
- Linux Foundation governance means MCP now has working groups, formal Spec Enhancement Proposals, and a roadmap driven by the people deploying it, not by one vendor.
Why this is a bigger deal than the hype suggests
For the past two years, every team building with AI has had the same problem. You want to give your model access to your CRM, your data warehouse, your ticketing system, your file storage, and your custom internal tools. You can either write a custom adapter for each one, or you can wait for a vendor to ship something. Neither option scales.
MCP solves the adapter problem by standardizing the wire format. A single MCP server for Salesforce works with Claude, GPT, Gemini, and every other model that speaks the protocol. You build it once, you use it everywhere.
The integration layer for AI was always going to get picked. Six months ago it was a fight. Today, MCP won.
What is actually painful about MCP right now
It is not all rosy. Teams running MCP at scale are running into the same set of problems, and the 2026 roadmap acknowledges every one of them.
- Auth is half-baked. SSO integration, scoped tokens, and per-user permissions are still being figured out at the spec level. Most enterprise deployments are duct-taping their own auth layer on top.
- Observability is thin. When an agent makes a decision via MCP, the audit trail is your problem to build. No standard format, no shared tooling.
- Gateway behavior is undefined. Running MCP behind a corporate proxy or API gateway works, but the patterns are still being established. Expect breakage.
- Configuration portability is rough. Moving an MCP server config from one environment to another is more manual than it should be.
- Governance is young. The Linux Foundation transition is recent. Decision-making processes are still being formalized.
A 5-step play if your team is starting from zero
- Pick one painful integration. The data source your model needs most. Salesforce, Postgres, Notion, whatever. That is your first MCP server.
- Build it as an MCP server, not a custom adapter. Even if you only have one model right now. Future-you will thank past-you.
- Wire your auth layer in front of it. Do not rely on the spec to handle this for you. Use your existing identity provider.
- Instrument every call. Log inputs, outputs, latency, and errors. Build the audit trail you wish the protocol gave you.
- Open source the server. If you are integrating a common system, someone else needs the same thing. Publish it. The community is hungry.
The strategic read
For most of 2025, the question for AI teams was “which model should we use?” In 2026, that question is mostly settled. The new question is “what does our model have access to?” MCP is the answer to that question, and it is the answer that won.
If your AI roadmap does not mention MCP yet, your roadmap is six months behind. That is fixable. Start with one server, ship it, then keep going.