WhyVault
Sign indeStart for free

Model Context Protocol

What is MCP? The USB-C port for AI assistants.

Suddenly everyone is talking about “MCP servers” and “connectors”. The idea behind them is simple: one shared plug that lets any AI assistant use any tool. Here’s how it works, no coding required.

Reading time: 9 minutesUpdated:

01

MCP in one picture: USB-C for AI

Remember when every phone came with its own charger? That’s roughly where AI assistants were until recently. If you wanted ChatGPT to check your calendar and Claude to read your notes, each pairing needed its own integration. Every vendor built its own plug.

MCP is the shared plug. Anthropic, the company behind Claude, released the Model Context Protocol as an open standard in late 2024. Since then OpenAI, Google, Microsoft and a long list of tools, from code editors to note apps, have adopted it.

What that means for you: a service that offers an MCP server works with any assistant that understands MCP. You pick the assistant you like, and your tools come along.

02

The four building blocks of MCP

That’s all the vocabulary you need to follow any MCP guide.

  • Client

    The AI assistant you talk to: Claude, ChatGPT, Cursor, Claude Code and many more. It decides when it needs a tool.

  • Server

    The program on the other end of the plug. It connects the assistant to a service: your notes, your calendar, your database.

  • Tools

    The actions a server offers, like search_notes or create_note. The assistant uses them like buttons.

  • Resources

    Content a server makes available to read, such as files or documents. Tools do something; resources are something.

03

Local or remote MCP server?

MCP servers come in two flavours. Which one you have decides where you can use it.

–Local MCP serverRemote MCP server
Where it runsAs a program on your computer, started by the appAt the provider, on the internet, at a URL
SetupAn entry in a JSON config file, often needing Node.js or PythonPaste the URL, sign in, approve
Sign-inUsually an API key in the configVia OAuth, like “Sign in with Google”, with no password in the assistant
In the browser & on your phoneNo, desktop apps onlyYes, wherever your assistant runs
Good forFiles and programs on that one computerServices you want to reach from anywhere
Where it runs
Local MCP serverAs a program on your computer, started by the app
Remote MCP serverAt the provider, on the internet, at a URL
Setup
Local MCP serverAn entry in a JSON config file, often needing Node.js or Python
Remote MCP serverPaste the URL, sign in, approve
Sign-in
Local MCP serverUsually an API key in the config
Remote MCP serverVia OAuth, like “Sign in with Google”, with no password in the assistant
In the browser & on your phone
Local MCP serverNo, desktop apps only
Remote MCP serverYes, wherever your assistant runs
Good for
Local MCP serverFiles and programs on that one computer
Remote MCP serverServices you want to reach from anywhere

A local server can do anything on your computer that you can. Only install servers when you know where they come from.

04

What is a connector in Claude and ChatGPT?

Connector is simply Claude’s word for a remote MCP server. Under Customize → Connectors you’ll find ready-made ones, for Google Drive for example, and you can add your own. All you need is the server’s URL.

ChatGPT calls the same thing an app and lets you add your own MCP servers through developer mode. In Cursor and Claude Code they’re just called MCP servers. Different names, same plug.

05

What MCP looks like in a conversation

You barely notice MCP. You ask a question and the assistant reaches for the right tool. Here’s Claude with the WhyVault MCP server:

Claude

  1. You: What do my notes say about the website relaunch?

  2. Tool: search_notes · “website relaunch”

    Claude: I found two notes. “Website relaunch” says: launch in November; still open are the copy for the pricing page and whether the blog moves over too.

  3. You: Add that the blog won’t come until January.

  4. Tool: append_to_note · “Website relaunch”

    Claude: Done. I added “Blog follows in January” under “Decisions”. You’ll see the change with my name in your history.

The grey lines are tool calls: which tool the assistant used, and with what.

06

What an agent can do with an MCP server

A server decides exactly which tools exist. Take WhyVault as an example:

  • Read

    vault_overview shows the structure, search_notes searches everything, read_note opens a note, get_backlinks shows what links to it.

  • Write

    create_note adds notes; append_to_note and update_section change single sections. Only with the read and write permission.

  • Not on the menu: delete

    There is no delete tool. If a tool doesn’t exist, the agent can’t call it. All tools

07

Is MCP safe?

MCP itself is just the plug. How safe it is depends on which server you connect and what permissions you give it. A server that only offers read tools can’t change anything. One with write tools can, and then you want to see what happened.

Also keep in mind: whatever your assistant reads through MCP, its provider sees too. When Claude fetches a note, Anthropic processes that text like any other message in the chat. So only connect services whose content you’d trust that assistant with anyway.

In WhyVault every change is saved as a version with the agent’s name on it, no agent can delete anything, and you can revoke any connection at any time. Connections that use an access token (Claude Code, Cursor) can also be set to read-only. More on history and permissions

08

How to spot a good MCP server

  • Clear origin: you know who runs the server and where your data lives.
  • OAuth sign-in instead of a password you paste somewhere.
  • Permissions per connection: revocable at any time; access tokens can also be read-only.
  • Targeted tools: edit a section instead of overwriting whole documents.
  • History with names: you see which agent changed what, and you can undo it.
  • No deleting by agents, or at least a trash you can restore from.
  • Your data can leave, for example as an export in an open format.

09

How to add an MCP server to Claude

Here’s how to connect a remote MCP server to Claude, using WhyVault as the example. It works on every Claude plan, including Free.

  1. 1

    Open Connectors

    In Claude, open Customize → Connectors, click +, then Add custom connector.

  2. 2

    Enter the URL

    Enter a name and the server’s URL. If Claude asks about the OAuth client, choose Register automatically. Click Add.

    Name: WhyVault
URL:  https://mcp.whyvault.io
  3. 3

    Sign in and approve

    Claude opens the service’s sign-in window. Sign in and approve what Claude is allowed to access.

  4. 4

    Turn it on in a chat

    In the chat, click + at the bottom left → Connectors and switch the connector on. Try: “Read my vault and summarise what I’m working on right now.” Full guide

10

MCP in ChatGPT, Cursor and Claude Code

ChatGPT: your own MCP servers go through developer mode under Settings → Security and login. It needs Plus, Pro, Business, Enterprise or Edu and only works on chatgpt.com. ChatGPT guide

Cursor: add the URL to ~/.cursor/mcp.json; the first time it connects, you sign in in the browser. Cursor guide

Claude Code: one terminal command does it: claude mcp add --transport http whyvault https://mcp.whyvault.io. Then type /mcp and sign in. Claude Code guide

Other apps like VS Code or Gemini CLI follow the same pattern. More clients

?

Frequently asked questions

What does MCP stand for?

MCP stands for Model Context Protocol. “Model” is the AI model, “context” is what it knows about your world, and “protocol” is the shared language the assistant and the tool use to talk to each other.

Do I need to know how to code?

Not for remote MCP servers: you paste a URL and sign in, like with any other app. Local servers take a bit more tech, because you edit a config file and often install Node.js or Python.

Is MCP safe?

MCP is just the plug. What matters is which server you trust and what permissions you give it. When in doubt, grant read-only access and prefer servers that version every change so you can undo it.

Does MCP cost anything?

No, MCP is an open standard and free to use. Costs can come from the assistant or the service behind the server. WhyVault, for example, is free up to 20 notes, and Pro costs €9.99 a month.

Does MCP work with free Claude?

Yes. On the free Claude plan you can add one custom connector; paid plans allow more. Claude guide

What about free ChatGPT?

No. ChatGPT only adds your own MCP servers through developer mode, which starts at Plus and only works in the browser. ChatGPT guide

What’s the difference between MCP and an API?

An API is an interface for programs, and every one looks different. MCP is a shared format in which services describe their tools so an AI assistant can understand and use them without extra code.

All guides

Plug in your memory.

WhyVault is a Markdown vault with a built-in MCP server. Connect once, and Claude, ChatGPT and Cursor read and write the same notes. Free, no credit card.

Start for free