What is an MCP server?

§ 01

What does MCP stand for?

MCP stands for Model Context Protocol. The official documentation describes it as an open-source standard for connecting AI applications to external systems: data sources such as files and databases, tools such as APIs, and reusable workflows. The same docs compare it to a USB-C port for AI applications: a shared way to plug systems into assistants, rather than a one-off integration for each product.

ChatGPT, Claude, and other MCP-compatible products can talk to the same kind of server, as long as that server is reachable and authentication is set up for that product.

§ 02

What can an MCP server do?

The official protocol docs group a server’s capabilities into three types. Most company work uses tools. The server can run on your computer (local) or on a host that assistants reach over the internet (remote). For company use, a remote server is the usual setup.

§ 03

How is an MCP server different from a built-in ChatGPT or Claude connector?

In ChatGPT, an app connects ChatGPT to an external service, such as Google Drive or Slack. A plugin packages capabilities into a workflow and can include one or more apps, skills, or both. Some plugins are only skills and do not connect to an outside system. When you need a system that is not already available as an app, OpenAI’s path is to add a custom MCP server and create it as a plugin.

In Claude, the connectors directory is a catalog of MCP servers that work across Claude products. You can also add a connector that is not in the directory by entering the URL of a remote MCP server.

Our honest take

MCP is the protocol. The server is the program. The ChatGPT app or Claude connector is the in-product wrapper. If a vendor already publishes one for your system, start there.

§ 04

When is a built-in connector enough?

A directory app or connector is usually enough when your system is listed, it covers the tasks you need, each person signs in with their own account, and you are comfortable with the vendor’s permission model. Official ChatGPT examples include Google Drive, Slack, SharePoint, Microsoft Teams, Outlook, Box, and Notion. Vendors such as HubSpot and Intuit QuickBooks also publish their own connectors; see how to connect QuickBooks, HubSpot or your CRM.

OpenAI is clear that app permissions do not grant new access; they only change when ChatGPT asks before using access the app already has. If you only need “search this Drive folder” or “look up this Slack thread,” a built-in connector is the simpler path.

§ 05

When does a company need a custom MCP server?

• An internal system, database, or API that no vendor lists in ChatGPT or Claude.
• A CRM, finance, or operations tool with no official connector, or one that is missing the records or actions you need.
• A workflow that must touch more than one system, with a limited set of tools rather than the vendor’s full catalog.
• A requirement that writes wait for approval, that only some roles can run certain tools, or that you keep your own audit log.
• A ChatGPT custom GPT that used custom actions, which do not transfer when the GPT is migrated to a plugin and may need a custom MCP server to replace them.

Our honest take

“Custom” means the server is built for your systems and your rules. It can stay private to your workspace. It does not have to be published in a public directory.

§ 06

How do you add a custom MCP server in ChatGPT or Claude?

The server has to be reachable by the assistant. For Claude, custom connectors connect from Anthropic’s cloud, and the server must be reachable over the public internet. For ChatGPT, OpenAI supports a public HTTPS endpoint or a Secure MCP Tunnel for private servers that should not be opened to the public internet.

• In ChatGPT, on the web: Plugins → plus button → Add custom MCP server. Enter a name, the server URL or a tunnel, and the authentication method (OAuth is the usual choice). Review the risk warning, create it as a plugin, and install it. Mention it with @ in a conversation. Write actions require confirmation by default.
• In Claude: Customize → Connectors, then add a custom connector by URL. On Team and Enterprise plans, an owner adds it for the organization; members then click Connect and sign in with their own account. Free plans are limited to one custom connector.

Our honest take

Done is a named person completing a named task, with the result matching the system of record, and with writes held for approval if that is your policy. The same server can often be added in other MCP-compatible products, such as Grok, but you should test the task in each assistant you use.

§ 07

What should “secure” mean for a company MCP connector?

OpenAI’s custom MCP guidance starts with a direct warning: only connect to MCP servers you trust. An untrusted server may access information shared through the app, or cause tools to run in ways you did not intend, including changing or deleting data. Claude’s docs give the same kind of caution for unverified custom connectors.

• People should sign in with OAuth or another method that does not require a shared password in a chat. MCP’s documentation recommends OAuth for remote servers.
• Each user should only get the access they already have in the source system.
• Tools should be scoped, with reads separated from creates and sends.
• Write and send actions should wait for approval. ChatGPT requires confirmation on writes by default.
• You should be able to turn tools off, disconnect the connector, and see a record of what was called.

Our honest take

A connector gives access to a system. A skill teaches the workflow. Most serious company setups need both.

§ FAQ

Questions we hear