What MCP means for influencer marketing agencies

What the Model Context Protocol is, what it changes for creator research, and the questions to ask before you connect any tool to Claude or Codex.

By Muatasim BaigUpdated 6 min read

The short answer

MCP, the Model Context Protocol, is an open standard that lets an AI app such as Claude or Codex use tools and data from other software. For an agency, it means you can ask Claude a research question and have it look the answer up in a data tool directly, instead of pasting exports into the chat. Before you connect anything, find out three things: what the tool can see, whether it can change or send anything, and how you check what it returns.

MCP in plain words

Anthropic open-sourced MCP in November 2024 as a standard way to connect AI assistants to the systems where data lives (Anthropic). The project's own documentation compares it to a USB-C port for AI applications: one standard plug instead of a custom cable for every device (MCP documentation).

There are three parts to keep straight:

  • The AI app you type into, such as Claude or Codex. MCP calls this the client.
  • The vendor's MCP server, which offers a set of tools, for example "search creators" or "look up a brand".
  • Your account with the vendor, which decides what those tools can reach.

Each tool has a name and a description, and the AI model decides when a tool would help with your question (MCP specification). You ask in plain words. The app picks the tool, calls it and shows you what came back.

Claude and Codex both support it. Claude supports custom connectors that point at a remote MCP server, on free and paid plans (Claude Help Center), and OpenAI's documentation covers adding MCP servers to Codex (OpenAI).

What it changes for creator research

Without MCP, research is a pile of tabs: a creator database, YouTube, a spreadsheet, and Claude off to the side for writing. Everything Claude knows is whatever you pasted in, frozen at the moment you copied it.

With a data tool connected, the question goes to the data. You can ask "Which brands sponsored home fitness creators this year, and in which videos?" and Claude calls the tool, reads the records and answers with them in front of it. A follow-up costs a sentence instead of a new export. And because the whole exchange stays in the chat, you keep a written trail of what you asked and what came back.

It doesn't fix everything. The model can still misread a record or summarize it badly. The tool's data has its own gaps and its own refresh schedule. A tool may also cap how many records come back per call, so a broad question can return a partial answer that reads like a complete one. Ask narrower questions, and ask Claude to say when it hit a limit. For a step-by-step research workflow, see How to use Claude for influencer marketing research.

Read-only versus write access

This is the question that matters most. Some MCP tools only read: search, look up, list. Others write: create a CRM record, change a status, send an email. Anthropic's help center says a server's tools can read data, create, modify or delete it, or take actions on your behalf, and that Claude can only reach what you've given that server permission to reach (Claude Help Center).

For research, read-only is what you want. A tool that can only look things up can't email a creator because Claude misread your request. The same help page warns that a malicious server can hide instructions meant to make Claude do things you didn't intend, which is one more reason to keep research tools to reading.

Two more details. The MCP specification says there should always be a human who can deny a tool call, and Claude shows you tool approval requests; Anthropic advises choosing "Allow always" only for servers you trust to run unsupervised. And a tool can label itself read-only, but the specification calls these labels hints, not guarantees, and says not to trust them from servers you don't trust (MCP specification). Ask the vendor rather than relying on the label.

What to ask a vendor before you connect

  1. Is the connection read-only? If any tool can write, what exactly can it change or send?
  2. Where does the data come from, and how fresh is it? Does each record carry a date?
  3. Does every answer link back to its evidence, such as the video or the post, so we can check it?
  4. How many records come back per answer, and what happens when there are more?
  5. How do we sign in, and how do we revoke access? In Claude you can remove a connector at any time under Customize, then Connectors.
  6. Which AI apps have you tested the connection with, end to end?
  7. What happens to the questions we send? Tool inputs go to your server, so who can see them, and how long do you keep them?
  8. What does an error look like? "No results" and "the tool failed" should never read the same.
  9. What does it cost, and is there a monthly limit on calls?

Run a first test in an afternoon

Don't roll a new connector out to the whole team on day one. Connect it on one account and ask it five questions you already know the answers to, from campaigns you've run. Check each answer against the evidence. When it's wrong, note why: missing data, stale data, or Claude misreading what came back. An afternoon of that tells you more about fit than any demo.

Questions

Do I need to be technical to use MCP?

No. In Claude you add a custom connector under Customize, then Connectors, by pasting the server address the vendor gives you and signing in (Claude Help Center). The vendor does the technical work.

Is MCP the same as an API?

Not quite. An API is how two programs talk, and an MCP server often sits on top of one. The difference is that the AI app finds the tools itself and decides when to call them, so you ask in plain words instead of writing code.

Does connecting a tool send my client's data to the vendor?

Whatever Claude sends as a tool input goes to that vendor's server, and the MCP specification recommends that apps show you those inputs before the call. Read the vendor's privacy terms, and leave client names out of your questions where you can.

Sources

  1. Introducing the Model Context Protocol, Anthropic
  2. What is the Model Context Protocol (MCP)?, Model Context Protocol
  3. Specification: Tools (version 2026-07-28), Model Context Protocol
  4. Specification: Schema reference, ToolAnnotations (version 2026-07-28), Model Context Protocol
  5. Get started with custom connectors using remote MCP, Claude Help Center (Anthropic)
  6. Model Context Protocol (Codex documentation), OpenAI

Each source was opened and checked on November 9, 2026.