Feature your MCP
/what-is-mcp

What Is MCP? Model Context Protocol Explained (2026)

What the Model Context Protocol is, what it's for and how to connect it to Claude, ChatGPT or Gemini. Explained without jargon, with real examples.

Nov 11, 2026· 5 min read

How MCP connects an assistant to a serviceDiagram: the AI assistant, which is the host, uses an MCP client. The client talks to an MCP server, and the server reaches a service or your data.AI assistanthostMCP clientMCP serverService or data

MCP (Model Context Protocol) is an open standard that lets an AI assistant connect to other apps and data sources to look things up or use them for you. The official documentation compares it to a USB-C port: one way to plug things in, whatever each of them does.

Without MCP, an assistant only knows what it learned in training and what you paste into the chat. With MCP it can read the current documentation of a programming library, check the latest runs of your GitHub workflows or drive a web browser for you.

MCP and Claude Fable 5.2

Anthropic released Claude Fable 5.2 on November 9, 2026, and Claude Code already uses it as its default Fable model. If you're wondering whether your MCP servers still work: they do.

MCP isn't tied to a model but to the app you use. The servers you connect in Claude or Claude Code are available to whichever model you choose in each conversation, whether that's Fable 5.2, Opus 5.5 or Sonnet 5.5. The same goes for ChatGPT or Gemini: you switch models and your connections stay.

What it's for: real examples

The easiest way to see it is with servers that already exist in the directory:

ServerWhat you ask your assistantWhat it does behind the scenes
Context7"Use the current documentation for this library"Uses resolve-library-id and query-docs to fetch up-to-date docs and code examples
GitHub"Show me the latest workflow runs in my repository"Uses its GitHub Actions tools to list workflow runs and fetch job logs
Playwright"Open this page and click the sign-up button"Drives a real browser: clicks, drags and uploads files on the page

In every case you just talk to the assistant. It decides when to use the tool, asks for your permission and shows you the result.

How it works

The diagram above sums it up. MCP has three pieces:

PieceWhat it isExample
App (host)The assistant you useClaude, ChatGPT, Cursor, VS Code
ClientThe part of the app that talks to each serverOne per connected server
MCP serverA program that gives access to something specific and explains what can be done with itContext7, GitHub

Under the hood they talk using JSON-RPC 2.0 messages, but you don't need to know that to use it, just as you don't need to know how USB works to plug in a mouse.

What a server can offer

TypeWhat it isExample
ToolsActions the assistant can runquery-docs, search, create a file
ResourcesInformation it can readDocuments, records, tables
PromptsReady-made instructions for common tasks"Summarise this pull request"

Almost everything you'll see in practice is tools.

Remote or local

RemoteLocal
Where it runsOn the internetOn your computer
How you add itYou paste a web addressYou edit a configuration file
What it needsSometimes signing in to the serviceUsually Node.js installed
Best forOnline services and public dataYour own files and programs

In Claude, remote servers are added as a custom connector (Add custom connector) and local ones in Claude Desktop's configuration file. In Claude Code a single command does it: claude mcp add. It's all step by step in how to install an MCP. Every Mcpteca listing tells you whether a server is remote or local, whether it needs a key and which assistants it works with.

Which assistants support it

The official documentation names Claude, ChatGPT, Visual Studio Code, Cursor and MCPJam, "among many others". Gemini CLI supports it too: you set it up with gemini mcp add or in its settings file. Not every plan or version allows the same things, so before connecting anything check whether your assistant shows an option for connectors or MCP servers. Coding agents in the terminal support it too: see our AI coding CLIs compared and the guide to AI agents. MCP also has a cousin for websites, explained in MCP vs WebMCP, and a companion format for know-how, the Agent Skills directory.

The current version: 2026-07-28

MCP doesn't use version numbers but dates. The current one is 2026-07-28. Its biggest change is that the protocol becomes stateless: the assistant no longer opens a "session" with the server before using it, which makes it easier to offer fast, cheap remote servers.

For people who just use MCP nothing visible changes. Anyone publishing a server should check the official changelog, because some changes break compatibility with earlier versions.

Who created it

Anthropic, the company behind Claude, introduced MCP in November 2024 as an open-source project. In 2025 other big players adopted it, starting with OpenAI, and in December of that year it moved to the Agentic AI Foundation, part of the Linux Foundation. Today it doesn't belong to any company.

Is it safe?

It depends on the server. The official specification is clear: a tool can take actions, so you should only connect servers from people you trust, and the assistant should ask for your permission before using them.

Before connecting one, check three things:

1. Who publishes it: that the author is who they say they are and has a recognisable repository or website. 2. What access it asks for: a documentation server doesn't need to read your email. 3. What you approve: read which tool the assistant is about to use before giving permission.

At Mcpteca we don't build or audit the servers, but every listing says where it comes from and which tools it offers when we've been able to check.

How to get started

1. Search the MCP server directory for something useful: public data, invoicing, programming… 2. Open its listing and check whether it's remote or local, whether it needs a key and which assistants it works with. 3. Connect it following the steps on the listing and try it with a simple question. Below there's an example you can set up right now.

Servers mentioned in this guide

Try it

/How to install

You need: Any Claude plan. The free plan allows 1 custom connector; on Team or Enterprise the organisation owner adds it first.

In claude.ai or the Claude app, open "Customize" → "Connectors".
Press "+ Add" and choose "Add custom connector".
Type a name you'll recognise, paste the URL below and press "Continue".
For authentication choose "No sign in", press "Continue" and then "Add".
https://mcp.context7.com/mcp

Afterwards: In a chat, press "+" at the bottom left → "Connectors" and turn it on.

You need: A paid ChatGPT account: it may not be available on the free plan. ChatGPT now calls connectors "Plugins".

Go to chatgpt.com/plugins.
Press "+" and choose "Add custom MCP server".
Type a name and a description, and paste the URL below under "Connection".
Don't set up authentication. Accept "I understand and want to continue" and press "Create as a plugin".
https://mcp.context7.com/mcp

Afterwards: In a new chat, type @ and pick the plugin.

You need: The Gemini app doesn't support custom connectors yet. It's done with Gemini CLI from the terminal, which needs a Gemini Code Assist license (it no longer has a free tier).

Install Gemini CLI if you don't have it yet.
In the terminal, run the command below.
Open Gemini CLI and type /mcp to check it shows up.
gemini mcp add --transport http -s user context7 'https://mcp.context7.com/mcp'

Afterwards: Ask as usual: Gemini will use the connector when it needs it.

You need: A Grok account. On Grok Business or Enterprise, an admin turns it on first.

Go to grok.com/connectors.
Press "New Connector" and choose "Custom".
Paste the URL below and save.
https://mcp.context7.com/mcp

Afterwards: Grok will use the connector whenever your question needs it.

Mcpteca does not review this server's code. A project can change hands or content: before you install or connect it, check its code or run it through a security reviewer.

Frequently asked questions

What does MCP stand for?

Model Context Protocol. It's the common way for an AI assistant to use external tools and data.

Do I have to pay to use MCP?

No. MCP is a free, open standard. Some servers or assistants have their own pricing; every Mcpteca listing says whether it needs an account or a key.

Does MCP work with Claude Fable 5.2?

Yes. MCP doesn't depend on the model but on the app: the servers you connect in Claude or Claude Code can be used by whichever model you pick, Fable 5.2 included.

Does it only work with Claude?

No. ChatGPT, Gemini CLI, Visual Studio Code, Cursor and many other apps support it too. What changes is where you add the server in each one.

Do I need to know how to code?

Not for a remote server: you just paste its address into your assistant. Local servers usually need Node.js installed and a configuration file edited.

How is it different from an API?

An API is a service's door for programmers. MCP is a common way for any assistant to use those doors without a custom-built connection for each one.

Is it safe to connect an MCP server?

It depends on the server. A tool can take actions on your behalf, so only connect servers from people you trust and check every permission your assistant asks for.

Who owns MCP?

Anthropic introduced it in November 2024 as open source. Since December 2025 it has been run by the Agentic AI Foundation, part of the Linux Foundation.

Browse the free directory of MCP servers, or tell us about yours.

Explore the directorySend us your MCP