What is MCP? A plain explanation for people who do not write code

MCP, short for Model Context Protocol, is an open standard that lets an AI assistant read from and act inside software it did not come with, without that software’s vendor building an AI feature first. Anthropic published it on 2024-11-25 (anthropic.com, 2024-11-25) and donated it in December 2025 to the Agentic AI Foundation under the Linux Foundation so that no single AI company controls it (blog.modelcontextprotocol.io, retrieved 2026-08-22). Its current published revision is 2026-07-28 (blog.modelcontextprotocol.io, retrieved 2026-08-22). For a person who does not write code, the practical effect is that the useful question about a work application has changed from whether its maker has shipped AI features into whether the AI you already have can reach it.

What does MCP actually do?

It gives every application a standard way to publish a list of things an AI assistant is allowed to do inside it, and gives every assistant a standard way to read that list. Each item on the list has a name, a description of what it does, and a statement of what information it needs to run. Your assistant reads the list, picks the item that matches what you asked for, fills in the details, and reports back what happened.

That is the whole mechanism. There is no intelligence in the protocol itself. The assistant supplies the judgment about which item to pick and what to put in it, the application supplies the actual capability, and MCP is the shared format that lets the two find each other without either one having been built with the other in mind.

You will see MCP compared to a USB-C port almost everywhere, including in the protocol’s own documentation (modelcontextprotocol.io, retrieved 2026-08-22). The comparison is fine for the standardization point and misleading on one thing worth knowing up front: a cable carries whatever you send down it, while an MCP connection carries only the specific actions the application chose to publish and that you were authorized for. It is closer to a library’s request slip than to a cable: you can ask for anything the catalog lists, in the way the slip allows, and nothing else. What is not in the catalog cannot be requested.

One housekeeping note, since the acronym is shared. MCP also stands for Master Control Program in the film Tron, for multichip package in hardware, and for management control panel in a few equipment manuals. This page is about the Model Context Protocol.

Do I need to know what an API is to follow this?

No. It helps to know that applications have always had a second, text-based entrance used by other software rather than by people, and that MCP is a standard shape for that entrance so an assistant can find it without being taught each one individually. That is the entire prerequisite.

The reason this is worth stating outright is that a common assumption gets in the way here. People tend to believe that AI capability sits on top of software expertise, so they assume they would need deep expertise in a tool before they could get an assistant to do anything useful in it. The dependency runs the other way in practice. What the assistant needs from you is four answers, and they are the same four used on every guide on this site.

1. What is this task’s one job?
2. What does a good result look like, specifically enough that someone else could tell?
3. What must never happen without my approval?
4. What supervision does this task need?

Those are answers about your work, not about the software’s internals, and you already have them.

The corollary is that a well-connected assistant asks you questions. If it asks which of two accounts to use, or which of three records is the right one, that is the mechanism working rather than failing.

Why would this exist if my apps could just add AI themselves?

Because the number of connections required grows as assistants multiplied by applications, and nobody is motivated to build that grid. Five assistants and twenty applications is one hundred separate integrations, each one maintained by a team that gets no benefit from making a competitor’s assistant work well inside its product. A shared standard turns those hundred into twenty-five: each application publishes once, each assistant reads once.

There is a second reason, less often said plainly. A product team’s incentive is to keep you inside its own interface, where it can show you its own features and its own upgrade prompts. Making a product easy to operate from outside runs against that incentive, which is part of why in-product AI features have arrived unevenly and why waiting for them is an open-ended wait. MCP changed the arithmetic without requiring any of those teams to change their minds.

Is MCP an Anthropic product, and could it be taken away?

It is not a product, and it is no longer Anthropic’s to withdraw. Anthropic published MCP on 2024-11-25 (anthropic.com, 2024-11-25) and donated it in December 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation, where it became a founding project. The foundation was co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. The same announcement records that MCP’s existing maintainers keep full autonomy over technical direction and that the Linux Foundation does not dictate it (blog.modelcontextprotocol.io, retrieved 2026-08-22).

This matters more to a non-technical reader than it looks. It is the difference between learning a feature and learning a standard. A feature belongs to one vendor and can be repriced, moved behind a higher plan, or retired. A standard under neutral stewardship is something the applications you use and the assistant you use both have to accommodate. Time spent understanding it is not a bet on one company’s roadmap.

Versions are published as dated revisions rather than numbered releases, which is a small thing that helps you read anything you find later. The revision current on 2026-08-22 is 2026-07-28, published on July 28, 2026 (blog.modelcontextprotocol.io, retrieved 2026-08-22). An article describing an older revision is not wrong, it is dated, and the date is in the name.

What is an MCP server, and do I need to run one?

An MCP server is the piece of software sitting on the application’s side that publishes that catalog of requests, and you almost certainly do not need to run one. The word server is the reason this term reads as intimidating, and it is a poor fit: in most cases the thing is either already running inside a service you use or is provided by a connector product, and the part you touch is a sign-in screen.

There are far more of them than most people expect. The protocol keeps a public registry at registry.modelcontextprotocol.io/v0/servers, a paginated open API whose limit parameter caps at 100 results. Walking all 243 pages with version=latest on 2026-08-22 returned 24,221 distinct server names. Counts near 7,100 circulate, including an earlier one on this site, and every one of them is a walk that stopped before the end.

The important caveat about that registry is who published each entry. Anyone can publish under their own namespace, so presence in the registry does not establish that the application’s own vendor made the server. Two checks settle it, and either one is enough. The first is a reverse-DNS namespace on the vendor’s own domain: a search for Linear on 2026-08-22 returns app.linear/linear, published on linear.app and carrying https://linear.app as its website, alongside five separate third-party Linear servers under names like io.github.pipeworx-io/linear and io.usefulapi/linear.

The second check is the one a domain comparison alone gets wrong. Many vendors publish under io.github.<organization>, which resolves to github.com rather than to the vendor’s own domain. io.github.github/github-mcp-server, io.github.firecrawl/firecrawl-mcp-server, io.github.SAP/fiori-mcp-server and io.github.getsentry/sentry-mcp are all the vendor’s own, and a domain-only rule would file every one of them as a stranger’s work. What settles those is the repository.url the registry holds for the entry, which names the GitHub organization that owns the product’s canonical code (registry queries, 2026-08-22).

So the useful habit is to read the prefix and then check the repository. This is the single most useful thing a non-technical reader can learn to do here, because a third-party server asking for access to your work account is a different decision from a first-party one. Which of forty-one widely used applications passed the check is on the best MCP servers page.

Does the AI I already pay for support MCP?

Several of the major assistants do, and the list has grown rather than fragmented. The protocol’s own documentation names Claude, ChatGPT, Visual Studio Code, Cursor and MCPJam as clients (modelcontextprotocol.io, retrieved 2026-08-22). The December 2025 governance announcement lists first-class client support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code as of that date (blog.modelcontextprotocol.io, retrieved 2026-08-22).

Support at the product level is not the same as availability on your particular plan or inside your particular organization’s account, and neither source above says anything about plan gates. Where it is available, it usually appears under a heading like connectors, apps, or tools rather than under the word MCP, which is why people conclude their assistant does not have it.

Can MCP reach the applications my company actually uses?

Some of them, and the honest answer for any specific application is to check rather than to assume either way. Reachability is uneven in a way the headline counts hide.

One catalog gives a sense of the shape. This site publishes 589 applications from Composio’s managed connector catalog, which holds 1,000 in total, with a median of 19 available actions each, queried 2026-08-22 from an inventory built 2026-08-14. The 1,000 is the ceiling of the tool that produced the snapshot, which records the truncation in its own metadata, so it is a floor rather than a count. The median is the number that matters more than the total. A typical connected application exposes enough actions for a bounded, specific job and not enough for an open-ended request, which is a fair description of what to expect on a first attempt.

Of those 589, 34 can start work on their own when something happens in the application rather than only when you ask, from the same query on 2026-08-22. That gap between 589 and 34 is the most practically useful number on this page. Almost everything reachable today is reachable when you initiate, and the fully automatic version of the idea is much rarer than the discussion around MCP suggests.

Where to check a specific application: the tools index has a page per application carrying its action surface, its in-product AI verdict, and the date each was checked. The four with the most demand behind them are Notion, Slack, HubSpot and Asana. For which applications have a server published by their own vendor rather than by a stranger, see the best MCP servers page.

Do I need IT to turn this on?

Often yes, in a managed work environment, and that is now a designed part of the standard rather than an obstacle to work around. An extension called Enterprise-Managed Authorization became stable on June 18, 2026. It lets an organization provision MCP access centrally through the identity provider it already uses, so a person gets their approved connections on first sign-in with no per-application authorization step, and it is described as being adopted by Anthropic, Microsoft and Okta (blog.modelcontextprotocol.io, retrieved 2026-08-22).

That announcement also names the reason so much of this is switched off in practice, in its own diagnosis of the previous model: “the per-user authorization tax keeps most of them switched off” (blog.modelcontextprotocol.io, retrieved 2026-08-22). The tools existed and the friction of authorizing them one at a time meant nobody did. If you have wondered why a capability you read about is not present in your account, that sentence is usually the answer, and it is an administrative fact rather than a technical limit.

Two questions are worth asking your administrator, and both are answerable without anyone learning the protocol. Which connections are already approved for my account. And what the process is for requesting one for a specific named application. Neither asks for a policy change.

This page stops at what one person can find out and ask for. Anything that promises a result across a whole team is a different kind of project with a different kind of budget, and it is not what a page like this can deliver.

The credential question is where a work setup differs most from a personal experiment, and it is worth understanding before you ask for anything. The informal route is an API key pasted into a configuration file, which then sits in that file indefinitely, carries whatever permissions it was issued with, and stops working without notice when it expires. A managed connector holds the authorization outside that file instead: the connector product defines the authentication method and the scopes that bound what the tools may do, and holds the credentials on its own side. Composio’s documentation describes this as an auth config that sets the method, the permitted scopes, and whose credentials are used, reusable across users (docs.composio.dev, retrieved 2026-08-22). The distinction is the difference between a secret you are personally responsible for and a connection an administrator can see, scope and revoke, which is also why the second kind is the one an IT team will say yes to.

What should I try first, and what is each option wrong for?

Start with something you can read rather than something you can change, because a read is checkable and cannot leave a mess. Beyond that, these are four different ways of working, each correct for some tasks and wrong for others. None of them is a stage on the way to another one.

The first row stays correct permanently for a large amount of work. Reading is where most of the real value of a connected assistant sits for most people, and treating it as something to be moved past is the most common way this goes wrong. The last row is the wrong choice for anything still in flux, because specifying a moving procedure produces a specification that is quietly incorrect.

What is MCP the wrong answer for?

It is the wrong answer whenever the problem is the work rather than the access. MCP does not add a capability an application does not have, so if the report you want does not exist in the tool, a connection will not produce it. It does not improve a procedure that is badly designed, and a connected assistant running a bad procedure runs it faster. It is a poor fit for high-volume repetitive entry that the application’s own interface already handles well with a bulk import.

Two further limits are worth being blunt about. A connection does not remove your obligation to check the result, particularly for anything that leaves your organization or moves money. And the fact that something is technically reachable is not a reason to reach it: plenty of work is reasonably done by hand, and the cost of specifying it exactly exceeds the cost of doing it.

What could not be established here?

• How many of the 589 catalog applications your specific company uses. No stack was examined for this page. The catalog figures describe what one connector product publishes, queried 2026-08-22 from an inventory built 2026-08-14, and say nothing about your account.
• Whether any of this saves time, or how much. No measurement was run. Any time saving attached to a connected assistant, including any figure produced later on this site, is an estimate until it is measured against the specific task.
• The true number of MCP servers in the public registry. Full pagination on 2026-08-22 returned 24,221 distinct names at version=latest. Servers exist outside the registry entirely, and an entry is a listing rather than a working service, so treat it as the size of the list and not of the field.
• Whether a given registry entry is the vendor’s own, beyond the ones named. Forty-one applications have been checked one at a time. The rest of the catalog has not, and a namespace check establishes who published a server rather than whether it is maintained, complete, or safe for a task.
• Plan and license availability. None of the sources cited here state which subscription tiers include MCP connections in Claude, ChatGPT, Gemini or Microsoft Copilot. Product-level support was verified; entitlement on your plan was not, and a feature named in vendor documentation is regularly absent from lower tiers.
• How many of the 589 a person can turn on unaided. The 589 published applications were not checked for how many are reachable without an administrator enabling something first. That count does not exist yet.
• How much of this reaches a managed work account. Enterprise-Managed Authorization is documented as stable and as being adopted by three named organizations as of June 18, 2026. Adoption by your identity provider was not checked.
• Demand figures for this topic. Google reports 14,800 monthly United States searches for what is mcp at a competition index of 5, and the AI channel returns 4,635 against 3,933 twelve months prior, measured 2026-08-22. The AI channel figure is modeled from People Also Ask data rather than observed assistant traffic, so read it as directional. The homonym exists and does not appear to contaminate the ranking: Google resolves the primary sense to Model Context Protocol and treats Master Control Program and the hardware senses as a footnote rather than a fork (SERP review, 2026-08-22). The AI-channel series is also the one funnel term on this site that has turned down recently: 3,933 twelve months ago, a peak of 5,309 in 2026-05, and 4,635 now. The year-over-year rise is real and the three-month direction is down.

Sources