What is MCP? The Model Context Protocol, explained plainly

Filed under Guide

By Gerald · 28 July 2026

Close-up of colorful patch cables connected across a bank of hardware ports, representing many tools wired into one system

Every explanation of MCP I read before writing this one starts with an analogy. USB-C for AI. A universal adapter. A translator between systems. Those lines are memorable and they explain nothing about what changes on your screen.

Here is the plain version. MCP, the Model Context Protocol, is an open standard that lets an AI assistant connect to your actual data and take actions in it, instead of you pasting text back and forth. You install or point the assistant at an MCP server, and from then on it can read your notes, your tasks, your files, or whatever that server exposes, and sometimes write to them too.

That is the whole idea. The rest is plumbing.

MCP is not a faster version of copy and paste. It is only worth the setup when the assistant needs to see data it genuinely cannot see otherwise.

What MCP actually is

MCP was introduced by Anthropic in November 2024 as an open specification, not a product you buy. Anyone can build an MCP server, and anyone can build an MCP client that talks to it. The protocol itself just defines the message format and the rules for a conversation between the two.

A server is the side that holds data or capability: your notes app, your calendar, your codebase, a search index. A client is the AI tool: Claude, ChatGPT, an editor, a custom agent. The client asks the server what it can do, then calls specific operations when the model decides they are useful.

The most important design decision in MCP is that it puts you, the model, and your data in one loop without a human manually relaying information between them. Before MCP, connecting an AI to your own systems meant either a custom integration built by the AI vendor, or you copying data into the chat window by hand. MCP standardizes the middle step so any client can talk to any server without a bespoke integration for each pair.

As of the 2026-07-28 specification release, the protocol has also settled some of its rougher edges. The stated changes, per the Model Context Protocol blog, include a stateless core so a server can scale on ordinary HTTP infrastructure without sticky sessions, an Extensions framework for optional capabilities, and two official extensions: MCP Apps, for server-rendered interfaces inside a client, and Tasks, for work that runs longer than one request and response. Authorization was also tightened to line up more closely with standard OAuth and OpenID Connect practice.

What an MCP server does that pasting text does not

A hand resting on a laptop keyboard at a desk in the evening, with a mug beside it — Once a connector is running, the AI reads your own data from inside the same window you already work in.

If your only need is "get this one paragraph from ChatGPT into my notes," MCP is overkill. Selecting text and pasting it is faster, and it will stay faster. I said the same thing plainly in why a connector beats copy-paste: copy-paste wins on speed, every time, for a single transfer.

MCP earns its cost in a different situation: when the value comes from the assistant seeing everything you already have, not from moving one thing. Say you ask an AI what you should work on this week. Without a connector, you are the one who has to remember which tasks are stuck, which notes are relevant, and which client conversation matters, then type all of that into the prompt by hand. With a connector, the model can read your actual task board, the checklist on a stalled item, and last month's project notes, then answer from what is really there instead of from what you remembered to mention.

That is the difference. MCP does not make one transfer faster. It removes the requirement that you manually reconstruct context every time you want a good answer.

Tools, resources, and prompts: what a connector exposes

An MCP server exposes its capability through three kinds of things, though servers rarely use all three.

Tools are actions the model can call, each with a name, a description, and a defined set of inputs. "Create a task," "search my notes," "move this card to done." The model decides when to call a tool based on the conversation, and a well-built server rejects calls that ask for something outside its rules, such as a delete action that was never implemented.

Resources are readable data the client can pull in directly, more like a file or a record than an action. A note's content, a document, a database row.

Prompts are reusable templates a server can offer, so a client can present a ready-made request instead of the user writing one from scratch. In practice, most consumer-facing MCP servers I have seen lean almost entirely on tools, because tools are what let the model act rather than just read.

The distinction matters for one reason: what a server does not expose, the model cannot do. A connector built with no delete tool literally cannot delete your data, no matter how the model is prompted. That boundary lives in the server code, not in a system prompt you have to trust.

What a live connector looks like on an ordinary Tuesday

Picture your notes, your task board, and your capture inbox all sitting in one app. You open Claude and ask what is actually overdue this week. The model calls a search tool, gets back your real task list with real due dates, and answers with the three things that are actually late, not three plausible guesses.

You mention a client call in passing. The model creates a task with the follow-up items, tags it to the right project, and links it to the meeting note it just read, because both of those already exist in the connected data and the tools to touch them exist too.

None of that required you to open a second window, find the right note, or type a summary of a project you already documented three weeks ago. The friction MCP removes is the step where you become the data-relay layer between yourself and the model.

Where MCP is still rough

I want to be honest about the parts that have not caught up. Setup still varies by client. Some tools need a URL and a token pasted into a settings page. Others expect a full JSON config file with server commands, which is fine for a developer and a real barrier for anyone else. That inconsistency is a client problem more than a protocol problem, but it is what you experience first.

Discovery is another gap. There is no polished directory where a non-technical user browses vetted MCP servers the way you would browse an app store. You mostly find servers by word of mouth or a vendor's own docs page.

And permission scoping differs a lot by implementation. A responsible server publishes exactly what it can do and refuses everything else. A careless one might expose more than a user expects, or fail to log what an agent actually called. Read what a server's tools can do, and what they explicitly cannot do, before connecting anything with write access to real data.

Who should bother and who should wait

Connect an MCP server if an AI assistant is already part of your daily workflow and you are tired of re-explaining your own context every session. If you plan a week from a task board, draft from your own notes regularly, or want an assistant that remembers what you have already decided without you repeating it, the setup pays for itself within a week.

Wait if you use AI for occasional one-off questions that do not depend on your personal data. Wait if the servers available for your tools are half-finished or undocumented; a bad connector is worse than no connector, because it either fails silently or does something you did not expect. And wait if you have not yet settled on which AI assistant you actually want to keep using, since reconnecting a server to a new client is extra setup you can skip until you know it is worth doing once.

Flow ships an MCP connector for exactly the case above: notes, tasks, and captures your AI can read and act on, without an AI subscription baked into Flow itself. The honest limitation is the same one every MCP tool shares. It is not a shortcut for a single copy-paste, and it will not feel worth it until you are actually asking the model to think with everything you have already written down. If you want the setup steps for specific tools, connect ChatGPT, Claude, and Perplexity to your notes and tasks walks through each one, and MCP servers worth connecting covers what else is worth wiring up beyond your own notes app.

Frequently asked questions

Is MCP an Anthropic product or an open standard? MCP is an open specification originally introduced by Anthropic in November 2024. It is not a paid product. Anyone can build a client or a server without a license from Anthropic, and the spec itself is maintained in the open at modelcontextprotocol.io.

Do I need to be a developer to use an MCP server? No, but setup difficulty depends entirely on the client. Some AI tools let you paste a URL into a settings page and you are done. Others expect a JSON configuration file, which is a real barrier if you are not comfortable editing config text. Check a specific tool's setup guide before assuming either way.

Does an MCP connector send all my data to the AI company? Only what the server exposes and the model actually requests during a conversation. A well-built server lets you audit exactly which tools exist and what each one can read or write, so nothing moves without a specific call. Read a server's published tool list before connecting anything with access to sensitive data.

What is the difference between an MCP server and a plugin? A plugin is usually built for one specific app or platform and does not work anywhere else. An MCP server follows an open protocol, so the same server can, in principle, work with any MCP-compatible client without the server author building a separate integration for each one.

Which AI assistants support MCP right now? As of mid-2026, Claude across its plans, ChatGPT on paid tiers with developer mode enabled, and Perplexity on paid plans all support connecting MCP servers, alongside a growing list of AI-first code editors. Support and setup steps differ by tool, so check the specific assistant's current documentation before assuming a feature exists.

My verdict

MCP is plumbing, and good plumbing is supposed to be boring. The only question worth asking before you set one up is whether an assistant answering from your real notes and tasks would actually change what it tells you. If yes, connect one. If you are just moving occasional text around, keep using copy and paste, it is faster and it always will be.