MCP Explained: The Protocol Anthropic Gave Away, and Why Everyone Uses It Now

If your AI assistant can now read your Google Drive, post to Slack, or pull a ticket out of Linear, there's a decent chance the plumbing behind that is the Model Context Protocol, or MCP. Anthropic introduced it in November 2024 as a fairly narrow piece of infrastructure — a standard way for an AI model to talk to outside tools and data. In December 2025, Anthropic donated it to a new Linux Foundation initiative, handing shared control to the same companies it competes with. That's an unusual move, and understanding why it happened says a lot about what MCP actually is and why it's worth knowing about even if you never write a line of code.

The problem MCP was built to solve

Before MCP, every AI product that wanted to connect to, say, GitHub or a database had to build a custom, one-off integration for it. Multiply that by every AI assistant and every tool, and you get an N×M mess: every model provider re-implementing the same connections over and over, with no shared standard. MCP's pitch was simple — define one protocol for "here's how a model asks a tool for information or asks it to do something," and let any AI product speak it to any tool that supports it. People have compared it to a USB-C port for AI: instead of a different cable for every device, one plug that everything agrees on.

How fast this actually moved

The adoption curve here is unusually fast for a technical standard. Anthropic shipped MCP in November 2024. By March 2025, OpenAI's Sam Altman announced support for it across the Agents SDK, ChatGPT's desktop app, and the Responses API — a notable moment, since OpenAI was adopting a standard built by a direct competitor rather than pushing its own. Google DeepMind's Demis Hassabis said Gemini would follow in April 2025, and Google shipped native MCP support in the Gemini API at I/O that May. Microsoft, GitHub, Cursor, and VS Code added support around the same stretch. According to Anthropic's own numbers, by the time of the December 2025 donation, MCP was seeing more than 97 million monthly SDK downloads and had roughly 10,000 active public servers, with client support baked into ChatGPT, Claude, Cursor, Gemini, Copilot, and VS Code.

Why Anthropic gave it away

In December 2025, Anthropic donated MCP's governance to the Agentic AI Foundation, a new directed fund under the Linux Foundation, co-founded alongside Block and OpenAI, with Google, Microsoft, AWS, Cloudflare, and Bloomberg also backing it. Anthropic said its own involvement wasn't changing — it's still investing in the protocol and maintaining core infrastructure — but the decision-making no longer sits with one company. That matters because a shared standard that's controlled by one of its competitors is a harder sell to everyone else building on top of it. Once Anthropic, OpenAI, Google, and Microsoft are all sitting on the same governance structure, no single vendor can quietly bend the protocol toward its own product. For something that's become this close to plumbing-level infrastructure, that kind of neutrality is arguably worth more to the ecosystem than Anthropic keeping the leverage of owning it outright.

What this actually changes for you

Most people won't ever touch MCP directly — it's a protocol, not a product. But it's the reason the "connectors" or "integrations" menu in whatever AI assistant you use has gotten so much longer over the past year without each one requiring a special app. If you use an assistant that can search your company's internal wiki, check a calendar, or query a database, someone stood up an MCP server for that data source, and your assistant is a client speaking a shared language to reach it. The practical upside is that adding a new capability to your AI setup increasingly looks like installing a browser extension rather than waiting for your vendor to build a bespoke integration.

Where to be careful

An MCP server is still a piece of software with access to your data or your accounts, and connecting your assistant to one is closer to installing a browser extension than downloading a PDF — it deserves the same scrutiny. Prefer servers published by the tool vendor itself (Google's, Slack's, GitHub's own) over unofficial community ones unless you've checked who maintains them, and pay attention to exactly what permissions a server is requesting before you connect it. The protocol's own 2026 roadmap acknowledges this is unfinished business: core maintainers have flagged audit trails, single sign-on, and more consistent gateway authentication as priority work for this year, specifically because enterprises adopting MCP at scale need better guarantees than "trust the server."

The bigger point

What's notable about MCP isn't the technical design — plenty of integration protocols have existed before it. It's that a standard built by one AI lab got adopted by its direct rivals fast enough, and became load-bearing enough, that the lab building it decided sole ownership was a liability rather than an advantage. That's a fairly rare sequence of events in software, and it's a decent signal that MCP is less a feature of any one AI product and more a permanent layer underneath all of them — worth knowing about the same way you'd know what HTTP does, even if you never write a request by hand.

Sources: Anthropic on donating MCP to the Agentic AI Foundation, Model Context Protocol blog on joining the Agentic AI Foundation, TechCrunch on OpenAI adopting MCP, TechCrunch on Google adopting MCP, Model Context Protocol 2026 roadmap.