What Is MCP? Why It Became the Connection Standard for AI Tools
We · October 2026
This is the first article in the "AI Tech Explainers" series, which takes the shape of one word, properly explained, per article: no product reviews, no rankings, no download or sign-up links. If the word "agent" is still vague, From Chatting to Getting Work Done is the place to start.
AI moves fast. The tools, capabilities and numbers in this article describe where things stand right now, and are for reference only.
First, the problem there was before any protocol
You have an AI application, and you want it to read your spreadsheets, query your database and open support tickets.
Before there was a protocol, every piece of software needed its own integration code: this one asks questions in a different shape from that one, returns results in a different shape, and you negotiate permissions with each of them separately. One application connecting to ten tools means ten integrations. Ten applications in the market, each needing those same ten tools, means the industry writes a hundred.
That is the phrase these protocols always lead with—the N×M problem: the number of integrations is the application count multiplied by the tool count. A protocol does exactly one thing: it turns multiplication into addition. Every application implements the protocol once, every tool implements it once, and now any pair can find each other.
What MCP actually is
MCP stands for Model Context Protocol. We do not lean on the full name much, because those three words do not help you use anything. Three other facts matter.
One: who proposed it, and when. Anthropic published and open-sourced it on 25 November 2024, with a metaphor that did more work than the specification: the USB-C of AI. It is not new electricity. It is a new socket.
Two: it splits one integration into three layers.
| Layer | In plain words | What it is on your machine |
|---|---|---|
| Host | The AI application you are actually using | A coding assistant, a desktop work agent |
| Client | The interpreter inside the host, holding one conversation with one endpoint at a time | Connection logic inside the app—you rarely see it |
| Server | The small program that actually holds the tools and the data | A process wired to your folders, database or ticket system |
The third layer is the important one: the tool side does not need to know who is using it. It can be this vendor's application or that vendor's, as long as both speak the same set of phrases. That is the physical reason your setup can be carried away.
Three: it defines three questions the model is allowed to ask. What can you do (the tool list), what do you hold (data and files), give me the result (the call and its return). The protocol governs manners, not intelligence—once something is plugged in, how well the job goes still depends on the model.
Why it counts as a standard
Whether a protocol becomes a standard is not decided by who stood on stage at launch. It is decided by who is still using it two years later. Here are the datable nodes:
| When | What happened |
|---|---|
| 2024-11-25 | Anthropic publishes and open-sources MCP, introducing the "USB-C of AI" framing |
| 2025-03 | OpenAI announces support in its Agents SDK; Google DeepMind, Microsoft (VS Code and Copilot), Cursor, AWS and Cloudflare follow |
| 2025-12-09 | Anthropic donates MCP to the Agentic AI Foundation, newly set up under the Linux Foundation; OpenAI and Block are co-founders, AWS, Google, Microsoft, Bloomberg and Cloudflare are platinum members |
| 2026-07-28 | The fifth revision of the spec ships, the largest ever: the core moves to stateless (sessions are removed) and adds extensions such as Tasks (long-running work) and MCP Apps (interactive UI) |
| During 2026 | ChatGPT, Gemini and M365 Copilot support MCP natively |
All of the above was checked in October 2026; the vendors' own pages are the authority.
The line that matters most is the December 2025 donation. A protocol started by one company gets handed to a neutral foundation, and two of that company's competitors sit among the co-founders. In software, that is usually the moment a standard stops being a product feature. It means no single company can decide where the road goes—which also means the configuration you wrote for one interface is unlikely to be voided because somebody changed their mind.
The other piece of evidence is on the product side: Chinese vendors broadly support the protocol directly—Zhipu's AutoClaw 2.0 adapts fully to MCP, WorkBuddy's Connector speaks both MCP and the command-line form, and Coze 3.0 can attach Claude Code and Codex CLI. Competitors, one interface standard. Historically, that agreement is a reliable signal. (For where each of these products sits, see the AI agent tool map.)
At the time of the donation, the publisher gave two scale numbers: roughly 97 million SDK downloads a month and over 10,000 active public servers. Those are the vendor's own figures, and we treat them as an order of magnitude rather than a measurement.
As for the July 2026 move to "stateless", the only conclusion ordinary users need is this: the interface started behaving like parcel delivery rather than a phone call. Nothing has to stay connected from start to finish any more, so a long job can be left running. What you feel is "I can hand it something that takes hours", not that a field in a specification changed.
MCP and skill packs are not the same thing
This pair gets confused constantly, so it deserves its own paragraph.
| Name | What it governs | An analogy |
|---|---|---|
| MCP | Connecting tools and data to the AI | The socket and the plumbing |
| Skill pack (SKILL.md) | Teaching the AI how to work your way | The instruction sheet taped next to the machine |
| Agent Plugins | Packaging the two above for distribution | A ready-made toolbox, closed and labelled |
Anthropic invented the skill-pack format in October 2025 and declared it an open standard; in October 2026 it donated that to the same foundation. Separately, OpenAI, Microsoft, AWS, GitHub, Cursor and Vercel are pushing a plugin packaging standard.
One sentence to keep the division straight: MCP decides what the AI can touch; skill packs decide what it does with it. Plug in as many endpoints as you like—if nobody taught it your process, it will still poke at your files the way the software's own buttons do.
Three practical consequences for you
First, what you wired up stays with you. The servers you installed all speak MCP, so switching to a different AI application should not mean rewiring them. We think this is the part of "standard" that matters to an individual: not the ten minutes of setup saved, but the cost of leaving comes down—you no longer keep using the app you started with just because seven connectors are already in place.
Second, every extra server costs you context. Endpoints are not free. The figure that circulates most widely in the field: hooking up five MCP servers can spend around 55,000 tokens just on telling the model which tools exist—the definitions alone (a media account of Anthropic's own documentation and engineering blog, checked October 2026). The protocol charges nothing, but the model re-reads those definitions on every call. So "I installed twenty servers" is not the same as "it works well"; more often it means every single request costs a little more. Article four of the agents series does that arithmetic properly.
Third, a tool endpoint is really a permission. A server that reads your files and makes requests for you can also send what it reads somewhere else. That is why the advice "only install servers whose maintainer you can identify" appears in every article about building your own tool chain. Capability and risk share one root; the full checklist is in article three of the agents series, and we will not repeat it here.
Three questions to size it up
Next time you read "supports MCP" in a product page, we suggest asking only three things:
- Is there an existing server for it? Maintained officially, or written by somebody in the community on a weekend?
- What does it need to read? Can you hand it only part of that?
- Does the application you use support the protocol natively, or does it need a translation layer bolted on?
If all three have clear answers, the connector is worth installing. If they do not, leave it off for now.
One more thing to keep straight while we are here: MCP hands the model the entrances to tools and data. Getting your own material in front of the model is a different chain — that is what What Is RAG: Letting a Model Answer With Your Documents in Hand is about.