What Is MCP? The Model Context Protocol, Explained

MCP — the Model Context Protocol — is an open standard for connecting AI models to tools and data. Anthropic introduced it in late 2024, and it solved a problem that was about to get ugly: every AI application needed custom integration code for every system it touched, and every integration had to be rebuilt for the next application.

Three things are worth knowing up front:

• It's a protocol, not a product. Nothing to buy. It defines how an AI application and a tool provider talk to each other.
• It won. OpenAI adopted it in March 2025, Google followed, and it now has first-class client support across Claude, ChatGPT, Cursor, Gemini, Microsoft Copilot, VS Code, and roughly twenty other clients.
• Anthropic no longer owns it. In December 2025 MCP was donated to the Agentic AI Foundation under the Linux Foundation, alongside Block's goose and OpenAI's AGENTS.md.

The rest of this covers the three roles that make it work, what changed in the current specification, and how you actually use a server.

The problem it solves

Before MCP, connecting an AI assistant to your company's ticketing system meant writing bespoke glue: authentication, a description of what the tool does, a way to pass arguments, a way to return results. Then you switched assistants and wrote it again.

The comparison people reach for is USB-C, and it holds: one connector standard means a device maker builds one port and every cable works. MCP means a tool provider builds one server and every compatible AI application can use it.

The scale of what followed says something about how real the problem was. Within a year, the protocol reached over 97 million monthly SDK downloads and 10,000 active servers, according to the MCP project's own write-up of its first year.

Host, client, server: the three roles

Almost all MCP confusion comes from these three words, so they're worth getting straight.

The host is the app. The server is the thing being connected. The client is the plumbing between them — one client per server connection, managed by the host.

You will see "client" used loosely to mean "the app," including by people who know better. When precision matters, the distinction between host and client is what decides whether a given server will work in a given application.

Servers expose three kinds of capability:

• Tools — actions the model can invoke, like reading a file or creating an issue
• Resources — data the model can read, like a document or a database record
• Prompts — reusable prompt templates the user can trigger

Tools are what most people mean when they talk about MCP. The other two are used less and matter for specific cases.

What the current specification changed

The specification is versioned by date, and the 2026-07-28 revision was a substantial restructuring rather than an increment. The headline changes:

A stateless core. MCP moved from a bidirectional stateful protocol to a request/response model. Sessions no longer need to be maintained, which means servers deploy behind ordinary load balancers like any other web service. That's a large practical difference for anyone running a server at scale.

Multi Round-Trip Requests. Replaces server-initiated requests. A tool can ask for user input mid-call without holding a stream open.

Header-based routing. Method and tool names travel in HTTP headers, so a gateway can route and authorize without parsing the request body.

Authorization hardening, including issuer validation and a shift away from dynamic client registration toward client metadata documents.

A formal extensions framework, so features can evolve outside the core specification.

The practical takeaway for someone choosing servers: a server that hasn't been updated since before this revision is working against an older shape of the protocol. Check the maintenance date.

Who governs it now

This matters more than governance usually does, because it's what makes the standard credible to companies that compete with Anthropic.

MCP was donated to the Agentic AI Foundation, a directed fund under the Linux Foundation, announced on 9 December 2025. The foundation launched with three inaugural projects — Anthropic's MCP, Block's goose, and OpenAI's AGENTS.md — and eight platinum members: Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI.

Anthropic's own announcement frames the original goal as an open standard rather than a proprietary one, and the donation is the version of that claim that's hard to argue with. A protocol controlled by one model vendor has a ceiling on adoption by its competitors. A foundation-governed one doesn't.

How to use an MCP server

The user-facing version is short, and it's configuration rather than programming.

1. Pick a server. Search the official registry for the system you want to connect — a database, a ticketing tool, your files. Prefer servers published by the vendor whose system they wrap over community wrappers, especially when credentials are involved. Our guide to the reference servers covers the official set and how to judge the rest.

2. Add it to your host. You provide either a command to run a local server, or a URL for a remote one, plus any credentials. Each host stores this in its own configuration; Anthropic's MCP documentation covers the Claude Code side, and the protocol's own site links the rest.

3. Restart the host. Most clients only discover servers at startup.

4. Know what your host does about approvals. Many hosts ask before a tool runs, and when yours does, read what it's about to do rather than clicking through. But don't treat the prompt as something the protocol guarantees. MCP doesn't specify a consent interface at all; whether you're asked, and when, is decided by the host and how it's configured. OpenAI's MCP tool, for instance, exposes an approval policy with always and never settings, and an agent running unattended may well be configured never to ask.

Which means the real security boundary sits earlier: connect servers you have reason to trust, give them the narrowest credentials and scopes that let them work, authenticate properly, and understand your host's approval policy before you rely on it. A dialog box you may or may not see is the last line, not the first.

Two things go wrong most often. A server that doesn't appear is usually a configuration file the host didn't reload, or a transport mismatch between what the server offers and what the host supports. And an agent that got noticeably worse after you added servers is usually a context problem: in the common setup, every tool definition from every connected server occupies space before your question does. Some hosts now load tool definitions on demand instead, so how much this costs you depends on which one you're using.

Where this leaves a non-developer

MCP is genuinely an infrastructure standard, and it reads that way. Setting up a server means editing a configuration file, handling credentials, and understanding transports — which is fine if that's your work and a wall if it isn't.

That wall is what Taku is built around: mirroring an AI setup that already runs into your own desktop workspace, so the configuration someone else got working becomes something you can use rather than something you have to reproduce. If you've read about connecting AI to your actual tools and never got past the setup step, the free app library is the short version. Taku is in Beta, and the Mac app is available now.

FAQ

What does MCP stand for?

Model Context Protocol. It's an open standard for connecting AI models to external tools and data sources.

Who created MCP and who owns it now?

Anthropic created and open-sourced it in November 2024. In December 2025 it was donated to the Agentic AI Foundation, a directed fund under the Linux Foundation, so it's now governed by a neutral body with backing from Anthropic, OpenAI, Block, Google, Microsoft, AWS, and others.

Is MCP only for Claude?

No. OpenAI adopted it in March 2025 and Google followed. It has first-class client support across Claude, ChatGPT, Cursor, Gemini, Microsoft Copilot, and VS Code, among others.

How do I use MCP servers?

Add the server to your host application's configuration — a command for a local server or a URL for a remote one — and restart the host. No programming required to use an existing server. Whether the host then asks before each tool call depends on its approval policy and your permission settings, not on MCP itself — so the protection that matters is connecting only servers you trust and giving them the narrowest credentials that work.

What's the difference between an MCP tool and a resource?

A tool is an action the model invokes, like creating a file. A resource is data the model reads, like a document's contents. Tools do things; resources provide context.

Does MCP send my data to Anthropic?

No. It's a protocol specification. Where your data goes depends entirely on which servers you connect and which model provider your host uses. A local filesystem server keeps file access on your machine.

Do I need MCP if I use Claude Skills?

They're independent and they compose. Skills carry instructions for how to do something; MCP carries connections to systems. Neither requires the other, and using both together is common.