What is MCP (the Model Context Protocol), and why does everyone suddenly care?
MCP, the Model Context Protocol, is an open standard for connecting AI models to tools, files, and apps. Think of it as a USB-C port for AI: one plug instead of a custom cable for every model-and-tool combination. Anthropic introduced it in late 2024. By 2026 OpenAI and Google support it too, and it's governed by the Linux Foundation.
For about two years, every “AI that connects to your tools” was a snowflake. Each app needed its own custom connector to each model. If you had five models and twenty tools, that was a hundred little bridges, and someone had to build and babysit all of them. MCP is the attempt to stop building snowflakes.
What is MCP, in one sentence?
The Model Context Protocol is an open standard that lets any AI model talk to any tool, file, or app through a single shared interface. The usual analogy is a USB-C port: before USB-C you had a drawer full of incompatible cables; after it, one plug fits everything. MCP is trying to be that one plug for AI.
Anthropic published it in November 2024. The point wasn’t to own it. They released it open, and the goal was for everyone to use the same plug.
What problem does it actually solve?
The technical name is the N×M problem. If you have N models and M tools and every pair needs its own custom connector, the work grows with N times M. That’s a lot of code that does nothing interesting and breaks whenever an API changes.
Teams reportedly spent most of their AI project time, with figures like 60–70% getting thrown around, just building and maintaining those connectors instead of doing anything useful with the AI. Treat that exact number as directional. The shape of the complaint is real even if the percentage is fuzzy. MCP collapses N×M into N+M: each model speaks MCP once, each tool speaks MCP once, and they all find each other.
How does it work, plainly?
Two pieces, two roles:
- An MCP server exposes a capability: your email, a database, a calendar, a folder of files. It’s the thing offering to do something.
- An MCP client lives inside an app you use, like an AI assistant or a code editor. It’s the thing that wants to use those capabilities.
They talk over a standard message format (JSON-RPC, if you want the technical term). Because the format is shared, a tool you wrap as an MCP server once can be used by any client that speaks MCP, not just the one you built it for. That portability is the whole pitch.
This is closely related to tool use and function calling, which an AI harness uses to let a model take actions. MCP doesn’t invent that idea; it standardizes the menu so the same tools work across harnesses.
Why does it matter now, in 2026?
A standard only matters if people actually adopt it, and that’s the part that changed:
- It’s no longer an Anthropic-only thing. OpenAI and Google DeepMind support it.
- There are hundreds of public MCP servers: pre-built connectors for common apps you can plug in rather than write.
- Since December 2025 it’s governed by the Linux Foundation, a neutral home, the way other internet standards are kept so no single company controls the spec.
When a standard gets a neutral steward and adoption from rivals, it stops being one vendor’s bet and starts being infrastructure. That’s roughly where MCP sits now.
What it changes for you
The honest version: MCP is the difference between an AI that knows things and one that can do things. A model with no connections can answer questions about your calendar only if you paste your calendar in. A model with an MCP connection to your calendar can read it live and add an event. The plug is what turns a clever box into something that actually touches the real world.
The caveat nobody puts on the box
A standard plug doesn’t make every appliance safe. Once you connect an AI to your real accounts, it can take real actions, and a wrong one isn’t a typo you catch later. It’s an email already sent or a record already changed.
So the plug is necessary but not sufficient. What the AI is allowed to do, and whether the steps it takes are trustworthy, matter at least as much as the connection itself. We get into that in why AI agents get things wrong.
Where this fits at Physea
We use MCP for the same reason everyone does: so our tools reach your apps through a mainstream port instead of bespoke glue. Cohesyn, our workspace for tasks, calendar, and people, speaks MCP, so an assistant can see your context through the same standard plug rather than a one-off integration we’d have to maintain forever. That’s not a feature to sell you. It’s just the sane way to build this now.
Common questions
- What does MCP stand for?
- Model Context Protocol. It's an open standard, first published by Anthropic in November 2024, for connecting AI models to external tools and data through one shared interface.
- Is MCP only for Anthropic's Claude?
- No. It started at Anthropic but OpenAI and Google DeepMind now support it, hundreds of public MCP servers exist, and since December 2025 it's stewarded by the Linux Foundation as a neutral standard.
- Do I need to understand MCP to use AI?
- Not really. It's plumbing. You benefit from it when an assistant can reach your calendar or email through a standard connection instead of a fragile one-off integration, but you rarely touch it directly.