A protocol is only as useful as the number of people who agree to use it. By that measure, the Model Context Protocol has had a remarkable fourteen months.

Anthropic released MCP in November 2024, framing it as an open-source answer to a mundane but serious problem: AI models cannot talk to the world's tools in any consistent way. Every integration was bespoke. A coding assistant that needed to read a file, call an API and write a result back to a database required three separate, hand-rolled connectors. MCP was meant to fix that by defining a single, standard interface — one socket that fits all plugs.

The comparison to USB is apt and has stuck. USB did not make computers faster or hard drives cheaper. It simply standardised the interface between them, and that alone proved transformative. MCP aims to do the same for the connection between AI models and the external systems they need to be useful.

From one lab's idea to everyone's problem

The protocol's early adoption was not guaranteed. Anthropic is a large company, but it is not the whole industry, and history is littered with open standards that arrived too early, too late or from the wrong sponsor.

What happened instead was unusually fast convergence. By March 2025, OpenAI had adopted MCP across its Agents SDK, Responses API and ChatGPT desktop application. That mattered enormously. The two largest players in the consumer AI space had agreed on a common interface before the standard had even celebrated its first birthday. A month later, Google DeepMind's chief executive, Demis Hassabis, confirmed that Gemini models would support MCP. The three dominant model families — Claude, GPT and Gemini — were, in effect, writing to the same socket.

For developers building agentic applications, this convergence was the signal they had been waiting for. The cost of building on a standard that might be abandoned or forked is the central risk in any nascent ecosystem. That risk shrank considerably when OpenAI and Google joined.

The spec grows up

Technical momentum alone does not make a protocol mature. In November 2025, a significant update to the MCP specification addressed the gaps that had accumulated as real-world deployments revealed what the original design had missed.

Asynchronous operations were formalised, allowing agents to kick off long-running tasks without blocking. Statelessness was better defined, making it easier to build MCP servers that scale horizontally. Server identity mechanisms arrived, giving clients a way to verify they were talking to whom they thought. Perhaps most practically, a community-driven registry began to take shape — a public index of MCP servers, searchable and maintained by the people who build with the protocol.

These additions are unglamorous. They are the kind of work that happens when a protocol moves from demos to deployment, from hackathons to production. Their presence in the spec is evidence that MCP's user base had grown large enough to care about edge cases.

Governance as the real announcement

Then, in December 2025, came the move that may prove most consequential of all. Anthropic donated MCP to the Agentic AI Foundation, a new body established under the Linux Foundation. The AAIF was co-founded by Anthropic, Block and OpenAI — a coalition that spans the protocol's creator, one of the most active commercial builders on top of it, and the company that had most recently adopted it.

Governance decisions in technology tend to be underreported and overimportant. When a single company controls an open standard, every other user of that standard is, in effect, a tenant. They can contribute, but they cannot own. The moment Anthropic transferred stewardship to a neutral foundation, the dynamic changed. Any company building on MCP now does so on ground that belongs, in principle, to all of them equally.

The Linux Foundation comparison is not accidental. Linux itself is an instructive precedent: a technology created by one person, donated to a foundation, and now the substrate on which much of the world's computing runs. The foundation model works because it aligns incentives. Companies that might otherwise compete on infrastructure can instead compete on what they build on top of it.

What 2026 is likely to look like

The consensus view among practitioners is that 2026 is when agentic workflows move from impressive demonstrations to daily practice. The tools exist. The models are capable enough. What remained was the connective tissue — a reliable, standardised way for an AI agent to pick up a task, reach into the systems it needs and set down a result.

MCP is that connective tissue. Or it is the leading candidate to become it, which for practical purposes amounts to the same thing. When a developer today wants to give an agent access to a database, a calendar, a code repository or a payment system, MCP is the obvious first choice. Not because it is perfect — it is not — but because it is where the ecosystem has gathered.

Standards win not by being the best option but by being the common one. USB was not, at launch, a technical triumph. It was a coordination triumph. MCP, fourteen months after its introduction, looks like both.