What is an MCP connector? Standard vs custom

An MCP connector is one authorized account behind an MCP server, exposed to your AI assistant as a set of tools. The word arrived with Claude and spread from there. If you use ChatGPT, Claude Code, Codex, Cursor or Antigravity you have probably seen the same thing called an MCP server; if you use Claude, Grok or Perplexity you have seen it called a connector. They are the same thing, and the vocabulary split is worth knowing before you go looking for one.

This post is about connectors and servers. If the protocol itself is new to you, start with What is MCP?

The short definition

An MCP server is a program that exposes data or actions to the app you ask your questions in (Claude, ChatGPT, Grok, Perplexity, Claude Code, Codex, Cursor or Antigravity) as a set of named tools. An MCP connector is one authorized account behind such a server: your Google Ads login, your Stripe key, your Search Console property.

The specification is stricter about the words than everyday use is, and it is worth borrowing its precision once: it calls that app the host, and reserves client for the component the host creates inside itself, one per server it connects to. So a host with four servers is running four clients. Nothing below turns on the distinction, but it is the reason documentation that says “client” sometimes means the app and sometimes means a connection.

The distinction matters because the relationship is one-to-many. One server can carry many connectors, and how many it carries turns out to decide what you can usefully ask it.

MCP stands for Model Context Protocol: an open specification for how the app and the server describe themselves to each other. The practical effect is that the model reads what the tools are at runtime, rather than someone writing an integration in advance.

How it differs from an API integration

Underneath, almost every MCP server calls an ordinary API. The difference is who reads the documentation.

	API integration	MCP connector
Who reads the docs	A developer, in advance	The model, at runtime
What gets written	Code that calls specific endpoints	Nothing: the server describes its own tools
Who decides what to call	The code, on a fixed path	The model, per question
What changes when the question changes	Someone edits the code	Nothing
Where credentials live	Your codebase or secret store	The server

That last row is the one to think about. An MCP server holds a credential and acts on your behalf, which is a real grant of access and worth evaluating properly. We wrote separately about what makes one safe to connect.

Standard and custom connectors

Claude sorts connectors two ways, and the other apps follow roughly the same split.

A standard connector is one the app lists in its own directory. In Claude that is the Connectors Directory, where each listing has a page setting out its use cases and what it can read and write, and connecting is a button.

A custom connector is any remote MCP server you add yourself by pasting its URL. Anthropic’s help page describes custom connectors as connecting Claude to services “that haven’t been verified by Anthropic”. That is the whole difference. The protocol is the same, the server is reached the same way, and the tools behave the same. What is missing is the listing.

Adding one is a settings step. In Claude: Customize, then Connectors, then Add, then Add custom connector. Name it, paste the server’s URL, and Claude walks you through signing in. Every plan can do this; Free is limited to one custom connector. On Team and Enterprise it is added once for the organization, under Organization settings, by an owner or primary owner (on Enterprise, also a role allowed to manage libraries), and members then connect it for themselves.

When you need a custom one

• The service is not in the directory yet. A server can be published in the official MCP Registry without being listed in a given app’s directory.
• Your company runs its own server. An internal tool will never be listed anywhere, and does not need to be.
• You use more than one app. A custom connector is a URL, and the same URL can be added to ChatGPT, Cursor or Codex in each one’s own settings.

HeyMetra is the first case: it is in the official MCP Registry and not in Claude’s directory, so in Claude you add it as a custom connector. The Claude setup page has the steps.

Anthropic’s advice on custom connectors is short: connect only to servers run by organizations you trust, read the permissions a server asks for, and let a tool run unsupervised only when you trust it to. We set out four questions worth asking before you connect one.

Why the connector count decides what you can ask

Most MCP servers are published by the company whose product they cover, which means most carry exactly one account. That is a perfectly good arrangement for questions about that account.

It stops working at the first question that crosses a boundary, and in marketing most of them do:

• What did this campaign cost, and what did the customers it brought in actually pay? The spend is in an ad platform, the revenue is in a store or a subscription platform.
• Which channel brings traffic that converts rather than traffic that bounces? Analytics on one side, orders on the other.
• Are the installs we paid for still subscribed three months later? Attribution on one side, subscription revenue on the other.

A server that sees one account can answer half of each of these. The half is not useful.

The obvious workaround is to add four single-account servers to the same app and let the model call each in turn. This does something, but the app is then doing the joining itself, from whatever fits in its context, without knowing that one system’s “conversion” and another’s “purchase” are the same event, and paying for the whole round trip in tokens each time you ask.

Holding the accounts behind one server moves that work to where the definitions are known. That is the entire design argument for a multi-account server, and it is the reason ours exists.

The vocabulary, since it is genuinely confusing

• Connector: Anthropic’s product name, and the word Perplexity puts on the same setting. If you are in either, this is what the settings screen says.
• MCP server: what everyone else calls it, including Anthropic once it is writing for developers rather than for the settings screen: its Claude Code documentation says MCP server throughout.
• Remote MCP server: one that runs on someone else’s infrastructure and is reached over HTTP. Nothing to install.
• Local MCP server: one that starts as a process on your own machine. The credential stays with you; so does keeping it running.
• Tool: one callable operation the server exposes. The set of tools is the whole of what a server can do.

Where to start

If you need one account, the vendor’s own server is usually the shortest path, including for accounts we support ourselves, because pretending otherwise would be obvious and pointless.

If the question you keep failing to answer is one that spans several accounts, that is the case for a server that carries several. Ours carries a catalog across ecommerce, CRM, subscriptions, app attribution and Search Console, and every connector page names the tools that account exposes.