What is an MCP connector? A plain guide for shop owners

What MCP is, in plain words#

An AI assistant on its own can write and reason, but it cannot see your customer list or press buttons in your email platform. The Model Context Protocol (MCP) is an open-source standard that gives AI applications one common way to reach outside systems: data sources, tools and workflows ^S1. The project compares it to a USB-C port: one shape of plug, many devices ^S1.

Anthropic introduced MCP in November 2024 as an open standard for two-way connections between AI tools and the systems where data lives ^S2. Today both Claude and ChatGPT support it ^S1, which means a service built once as an MCP server can be used from either assistant.

An MCP connector is what you, the shop owner, actually add: a named entry in Claude or ChatGPT that points at an MCP server and signs you in to it. After that, you can ask in plain words and the assistant decides which of the service’s tools to use.

Server, client and connector#

Three words get mixed up in most explanations. The protocol itself defines a host, clients and servers ^S3:

Claude calls these custom connectors and lets you connect existing remote MCP servers or build your own ^S4. In ChatGPT, you add one from Plugins with “Add custom MCP server”, then save it as a plugin ^S5. The names differ; the protocol underneath is the same.

Local and remote MCP servers#

MCP servers run in one of two places ^S3:

• Local servers run on your own computer and talk to the AI app through standard input and output. They suit developer tools and files on your machine, and usually serve one user.
• Remote servers run on the provider’s platform and are reached over the internet with the Streamable HTTP transport. They serve many users, and the protocol recommends OAuth for signing in ^S3.

For marketing, remote is the practical choice. Your contacts, sending domain and campaign history live with the service, nothing needs installing, and the same connector works from your laptop, your phone and a teammate’s account.

Tools and what a connector is allowed to do#

A server offers three kinds of building blocks: tools (actions the AI can call), resources (information it can read) and prompts (reusable templates) ^S3. Tools matter most, because they are where things happen: search contacts, create a segment, render a preview, schedule a send.

Tools are designed to be called by the model when your request needs them ^S6. Each tool can carry annotations, such as whether it only reads data or can change or delete it, so the AI app can treat risky calls differently. The specification also warns that apps should not trust those labels unless the server itself is trusted ^S6.

What a connector can reach depends on how you sign in. For remote servers, MCP builds its authorization on OAuth 2.1: you approve access in the browser, and the access token is issued for that one server and must be refused by any other ^S7. A server can also reject a request when the granted scopes do not cover it ^S7.

What makes a connector safe for actions with consequences#

Reading your list is low risk. Emailing 3,000 customers is not: a campaign cannot be recalled once it lands. The MCP specification says there should always be a human in the loop who can deny a tool call, and that apps should ask for confirmation on sensitive operations ^S6. Claude’s help article adds practical advice: connect only servers from organizations you trust, review what a server requests when you sign in, and choose “Allow always” only for tools you trust to run unsupervised ^S4.

Those are client-side controls. A well-built marketing connector adds its own safeguards on the server, so safety does not depend on a setting someone forgot to change. Look for these:

• A separate approval step. Preparing a send and confirming it should be different steps, with the confirmation done by a person. An approval card in the chat is one way to do this.
• A frozen audience. What you approve should be exactly what goes out: the recipient count and exclusions fixed at the moment you confirm.
• Narrow sign-in. Read and draft access through OAuth, with no API key that can send on its own.
• A way back. A cancel window before a scheduled campaign goes out.
• Guardrails on volume. Automatic pauses when complaints or bounces rise.

Marketing tasks an MCP connector can handle#

For a small apparel brand such as Ninefold, which sells knitwear, outerwear and basics, a marketing connector turns a request in plain words into the steps a dashboard would otherwise make you click through:

• “How many customers bought outerwear last winter but nothing since?” The assistant sizes a segment before you save it.
• “Draft an email introducing the new merino crewnecks, in our brand style.” The connector supplies the brand, the assistant writes the copy.
• “Show me the mobile version and shorten the headline.” An interactive preview updates in the chat.
• “Send me a test, then schedule it for Thursday at 9 am.” A test goes to you; the send waits for your approval.
• “How did last week’s campaign do compared with the one before?” Results come back in the same conversation.

Some servers can show interactive cards instead of plain text. MCP Apps is an extension that lets a server return an interface, such as a form or a preview, that renders in a sandboxed frame inside the chat ^S8. Support varies by client ^S8. For ideas on prompts, see writing marketing emails with AI.

How Tellnod works as an MCP connector#

Tellnod is a remote email marketing MCP server. You add one address in Claude or ChatGPT, sign in with Google, and keep your contacts, brand, sending domain and results in the chat you already use. Setup takes about a minute on the connect page; the tool list, sign-in details and limits are on the MCP server reference.

Sending uses a two-step flow. The assistant can prepare a campaign, which freezes the recipient list and shows a confirmation card with the recipient count, who is excluded and why, and the send time. Only the Confirm button on that card can schedule the send; the assistant cannot call it. Each confirmation is single-use and expires after 30 minutes, and a scheduled campaign can be cancelled until it goes out.

Email works today. WhatsApp and LINE take early-access requests and are designed to use the same contacts and the same approval. Step-by-step guides: send a campaign from Claude and send from ChatGPT. Plans start free for up to 500 contacts; see pricing.

Questions to ask before you add a marketing connector#

1. Who builds and hosts the server, and do you trust them with customer data?
2. What does sign-in ask for, and can you limit it?
3. Which tools only read, and which change data or contact customers?
4. Can the AI complete a send by itself, or does a person have to approve it?
5. Can you see the real message and the audience before approving?
6. Is there a cancel window and an automatic pause on complaints?

FAQ#