Claude MCP: How to Connect Claude to Your Marketing Tools
How Claude connects to outside tools through MCP: the three surfaces, local and remote servers, setup steps from the docs, permissions and security mistakes.
02 / In this article
Claude MCP means using the Model Context Protocol to let Claude read from and act in tools outside the chat, such as an ad account, an analytics property or a CRM. As of October 2026 you can do it in three places: as a connector in Claude and Claude Desktop, as a configured server in Claude Code, and through the MCP connector of the Claude API. The setup is short. The decisions that matter come after it: which server you trust, which permissions it gets and who approves each action.
Key Takeaways
- Claude reaches tools through MCP servers. A server exposes tools, and Claude calls them when your request maps to what a tool describes.
- There are three surfaces. Connectors in Claude and Claude Desktop,
claude mcpin Claude Code, and the MCP connector in the Messages API. They differ in who sets them up and where the server runs. - Remote and local are different risks. A remote server runs somewhere else and signs you in with OAuth. A local server is a program that runs on your machine with your permissions.
- Anthropic does not verify custom connectors. The help center says to connect only servers from organizations you trust and to review tool approval requests.
- Start read-only and keep a human on every change. Reporting is a safe first use. Edits to budgets, campaigns or tracking should always be approved by a person.
What is Claude MCP?
Claude MCP is the use of the Model Context Protocol, an open standard for connecting AI applications to external systems, so that Claude can call the tools those systems expose. The protocol has two sides. The MCP client is built into the Claude application you use. The MCP server is a program or a web service that offers a list of tools, for example "list campaigns" or "run a report".
When you ask a question, Claude decides whether a tool fits. Anthropic's API documentation describes the rule: Claude calls a connected tool when the request maps to the tool's described capability, and it answers general questions directly. Asking how a database product works gets an answer from the model. Asking what is in your own database triggers the tool. If you want the protocol itself explained from the beginning, our guide to what MCP is covers the architecture and the vocabulary.
The practical effect for a marketer is that the data stays where it lives. You do not export a CSV and paste it into a chat. Claude asks the source, with the permissions you gave it, and works on what comes back.
Where can you use MCP with Claude?
You can use MCP with Claude on three surfaces, and each one has a different setup path. Pick the surface first, because it decides who configures the server and where it runs.
Claude and Claude Desktop
In the Claude apps, MCP shows up as connectors. Anthropic's help center describes two kinds: directory connectors, which are prebuilt and listed with their read and write capabilities, and custom connectors, which point to any remote MCP server. According to the help center, custom remote connectors are available on Claude, Cowork and Claude Desktop for Free, Pro, Max, Team and Enterprise plans, and Free users are limited to one custom connector. On Team and Enterprise plans an Owner must enable a connector for the organization first, and then each person authenticates individually.
Claude Desktop can also run local servers. The MCP project documents this with a configuration file, and we walk through it below.
Claude Code
Claude Code is Anthropic's command line coding tool. Its documentation describes adding MCP servers with the claude mcp add command, and storing them at three scopes: local, project and user. It suits technical users, and a shared .mcp.json gives a team the same servers.
The Claude API
The MCP connector in the Messages API lets an application connect Claude to remote MCP servers without writing an MCP client. The documentation marks it as a beta feature. It supports tool calls only, and the server must be publicly reachable over HTTP. Local stdio servers cannot be connected directly. This is the surface for someone building their own product or automation on top of Claude, not for a person chatting.
How do remote and local MCP servers differ?
A remote MCP server runs on someone else's infrastructure and is reached over HTTP, while a local MCP server is a program that starts on your own computer and talks to the app through standard input and output. The difference decides how you sign in, where your data travels and what the server can touch.
| Remote server | Local server | |
|---|---|---|
| Where it runs | The provider's infrastructure | Your computer |
| How Claude reaches it | Over HTTP | Started as a process, stdio |
| Typical sign-in | OAuth in a browser, or a header | Environment variables or files on your machine |
| What it can touch | What the service's API allows for your account | Whatever your user account can do on the machine |
| Main risk | Over-broad permissions, unverified provider | Running unknown code with your rights |
For remote connectors, Claude connects to your server from Anthropic's cloud, not from your device. A server on a private network is not reachable unless Anthropic's IP addresses are allowed.
For local servers, the MCP documentation is direct: the server runs with your user account permissions, so it can perform any file operation you can perform manually. That is why the next section starts with the lightest possible setup and why the later sections spend time on permissions.
Want Claude connected to your marketing stack?
Tell us which tools hold your data and which tasks eat your week. We reply with what could be automated and what should stay manual.
Explore AI workflowsHow do you connect Claude to a tool step by step?
The steps depend on the surface. The ones below are taken from the official documentation as it stands in October 2026, with the page named each time. Interfaces change, so check the page before you follow a step.
How do you add a custom connector in Claude?
Anthropic's help center article on custom connectors lists the steps for Pro and Max accounts. Go to Customize, then Connectors, click "Add" and choose "Add custom connector". Enter a name and the remote MCP server URL, review the detected authentication settings, and choose how to sign in. Options include "Sign in now", "Sign in when needed" and "No sign in". If the server does not use OAuth, you can add fixed credentials such as an API key under request headers. Finish with "Add".
On Team and Enterprise plans, the steps move to the organization. Per the same article, an Owner adds the connector under Organization settings, then members open Customize, find the connector, and click "Connect". Each user signs in on their own, so Claude only accesses what that user can access. The exception is a connector built on a shared credential such as an API key, which gives everyone the same access.
To use a connector in a chat, click the plus button at the lower left of the conversation and open "Connectors". Each one has a toggle for that conversation. Turn on only what the task needs.
How do you add a local server to Claude Desktop?
The MCP project's guide to local servers uses Claude Desktop and the Filesystem server as the example. In the Claude menu, open Settings, then the Developer tab, and click "Edit Config". The file is ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows. The guide gives this structure for the macOS example:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/username/Desktop",
"/Users/username/Downloads"
]
}
}
}
You replace username with your own, list only the folders the server may use, save the file, and fully quit and restart Claude Desktop. The guide notes that all actions require your approval before they run, and it warns to grant access only to directories you are comfortable with Claude reading and modifying.
The command is what runs on your computer, so read it before you save it.
How do you add an MCP server in Claude Code?
The Claude Code documentation on MCP gives the command forms. For a remote server over HTTP, which it lists as the recommended transport for remote servers:
claude mcp add --transport http <name> <url>
For a local stdio server, the options go before the name and the server command follows a double dash:
claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable \
-- npx -y airtable-mcp-server
The same page documents claude mcp list to see what is configured, claude mcp remove <name> to remove a server, and the /mcp command inside Claude Code to check status and sign in to servers that use OAuth. Its status list includes "Connected", "Needs authentication", "Failed to connect" and "Pending approval".
Three scopes decide who gets the server. Local, the default, loads only in the current project for you and is stored in your ~/.claude.json. Project scope writes a .mcp.json file at the project root so a team shares it through version control. User scope loads the server in all your projects. For project-scoped servers, interactive sessions prompt you to approve them before use, which matters when you clone a repository you did not write.
How does the API connector work?
In the Messages API you pass the server in an mcp_servers list and add a matching mcp_toolset entry in tools. The feature is in beta and requires a beta header, which the documentation names mcp-client-2025-11-20. Copy the current request shape from the documentation page, because beta parameters change. The page also describes allowlisting and denying specific tools.
What can marketers actually do with Claude MCP?
Marketers use Claude MCP to ask questions of live data and to hand repetitive reading and drafting to Claude, while a person keeps the decisions. The examples below are generic. They describe what a connection makes possible, not results anyone should expect.
Ad account review. With a server for an ad platform, you can ask for a summary of spend and results by campaign for a date range, or ask which ad sets have no impressions. Reading and summarizing is the safe start. Changing bids or budgets is a different class of action, covered below. For a deeper look at one platform, see our article on Google Ads MCP.
Analytics questions. A connection to an analytics property lets you ask in plain language which pages brought sessions last month or how a channel compares with the previous period, then ask a follow-up without rebuilding a report. The answer is only as good as the tracking behind it, which is why tracking setup comes first.
CRM hygiene. A connection to a CRM lets Claude look up a contact, summarize the history of a deal or list records missing a field. Writing to the CRM is where you want approvals on, because a wrong bulk update is slow to undo.
Reporting drafts. Combine two or three read-only connections and ask for a first draft of a weekly summary. A person edits it. Our guide to MCP tools for marketers compares the kinds of servers available.
The value comes from removing copy and paste between systems. The judgment about what to do next does not move.
How do permissions and approvals work?
Permissions work in layers: the service decides what your account may do, the connector or server decides which tools exist, and the Claude app asks you before a tool runs. Each layer can only narrow the one above it.
The help center states the first layer plainly: Claude inherits each person's permissions from the connected service, so it cannot see anything you cannot already access there. For Team and Enterprise plans, Owners can set tool permissions per category or per tool to "Always allow", "Needs approval" or "Blocked", and those settings apply across the organization. The article's own example is allowing Claude to read files in Google Drive while blocking it from creating or editing documents. It adds that these restrictions only narrow access, and a person still needs the underlying permission in the source system.
The second layer is what you do in the chat. The help center advises reviewing tool approval requests carefully and choosing "Allow always" only for servers and tools you trust to run unsupervised. It also points to the "Search and tools" menu, where you can disable tools that are not relevant. In Claude Code, project servers need your approval, and the /mcp command lets you check and toggle servers. In the API, you choose which tools to enable and which to deny, so a coding mistake on your side can grant more than you meant.
One caution from the help center deserves its own line. With Research, Claude can call connector tools without further approval, so Anthropic advises disabling any tools that can write to external apps.
What is a safe default?
A read-only connection tested with a question whose answer you already know. Switch any tool that can change something to "Needs approval" or block it until you have a reason to allow it. If the service lets you create a read-only credential, use it instead of your own login.
Who approves what in your workflows?
We map which steps can run on their own and which need a person, then design the checks around them.
See business process automationWhat are the most common security mistakes?
The most common mistakes are trusting an unverified server, storing secrets in plain text, granting more access than the task needs, and letting write actions run without a human. All four are named in the vendors' own guidance.
Connecting a server nobody has vetted
Anthropic's help center says custom connectors use servers it has not verified, and that you should only connect to trusted organizations and review the permissions requested at sign-in. It adds that server developers may change tool behavior without notice. A server that was fine last month can change. Prefer servers published by the service provider itself, and look for a public repository or documentation you can read. Anthropic also provides a route to report a malicious MCP server through its vulnerability disclosure program.
For local servers, the MCP security guidance names the risks: a malicious startup command in a configuration, a malicious payload inside a server, or an insecure server left running. Consequences include arbitrary code execution, data exfiltration and data loss. Read any command before you paste it into a configuration, including the ones in this article.
Prompt injection through tool results
Anthropic's help center and the Claude Code documentation both warn about prompt injection. A server that fetches external content can return text containing instructions aimed at the model. Be careful when a task combines a source you do not control, such as inbound email or web pages, with a tool that can write or send.
Secrets in plain text
Claude Code lets you put credentials in a project's .mcp.json, which is meant to be committed. The documentation supports environment variable expansion in that file, with the form ${VAR} and a default form ${VAR:-default}, so the secret stays in your environment and the file holds only a reference. Use that for tokens and API keys. Keep them out of chat messages, shared documents and repositories. If a key is ever pasted somewhere it should not be, rotate it.
Rights that are too broad
The MCP security guidance has a section on scope minimization. It warns that broad scopes granted up front enlarge the damage if a token leaks, make revocation more disruptive and blur the audit trail. It recommends starting with a minimal set of low-risk read operations and raising access only when a privileged operation is needed. As a user, apply the same idea: grant the smallest scope the sign-in screen allows, and use a dedicated read-only credential where the service offers one.
Letting writes run unsupervised
Ads, budgets and tracking are where an unreviewed change costs money or corrupts data. A person should review any change to ad accounts, budgets, audiences or tracking tags, and write tools you do not need should be disabled.
Forgetting to disconnect
The help center recommends disconnecting services you no longer use. In Claude Code, claude mcp remove also deletes a remote server's stored OAuth tokens.
What if no server exists for your tool?
If no existing server covers your tool, a developer can write one with the official SDKs. Our guide to building an MCP server covers the first steps. The security rules apply from the other side too: the MCP guidance states that servers must not accept tokens that were not issued for them.
Claude MCP is a connection layer, not a strategy. Start from the question the business needs answered, find the data source, and only then decide whether a connection helps. That is the order we use in our AI workflow work, and in a growth strategy, where measurement and priorities come before tools. Claude is one of the AI tools we work with, alongside others such as ChatGPT: see ChatGPT MCP. For the wider picture, read AI in digital marketing.
A small first project: pick one weekly report, connect the one read-only source it needs, and compare Claude's draft with your own for a few weeks.
Frequently Asked Questions
What does MCP stand for in Claude?
MCP stands for Model Context Protocol. It is an open standard that lets AI applications such as Claude connect to external tools and data through servers. The server lists the tools it offers, and Claude calls them when a request fits.
Is Claude MCP free to use?
It depends on the surface and the plan. According to Anthropic's help center in October 2026, custom remote connectors are available on Free, Pro, Max, Team and Enterprise plans, with Free limited to one custom connector. The services you connect may have their own costs or API limits. Check the current pages for your plan.
What is the difference between Claude Desktop MCP and Claude Code MCP?
Claude Desktop is a chat application where you add connectors in the interface or edit a configuration file for local servers. Claude Code is a command line tool where you add servers with claude mcp add at local, project or user scope. The first suits conversations with your tools, the second suits work inside code and project folders.
Do I need to code to connect Claude to a tool?
Not for remote connectors in Claude: you add a name and URL and sign in. Local servers need a configuration file and often Node.js. Claude Code and the API need command line or code skills.
Can Claude change things in my tools?
That depends on the tools the server exposes and the account you signed in with. You can disable tools, require approval, or use a read-only credential.
Is it safe to connect Claude to my ad account?
It can be done carefully, but it is not risk free. Anthropic states that custom connectors are not verified and that Claude can take actions in connected services based on your permissions. Use read-only access first, review every approval prompt, and have a person approve any change to budgets or campaigns.
How do I remove or revoke a connection?
In the Claude apps, disconnect the connector in the connector settings. In Claude Code, run claude mcp remove <name>, and use "Clear authentication" in the /mcp menu to revoke access. Also revoke the token or API key in the connected service itself if you want to be sure.
Next steps
Choose one read-only task, one data source and one surface, and test it on a question you can verify. Keep a person on every change. To decide which steps are worth connecting, look at our AI workflows service and our business process automation page, or read how measurement shapes the plan in analytics and tracking. For more on the protocol and the tools around it, start with what MCP is and the MCP tools for marketers overview. When you are ready to talk it through, get in touch.
Sources
- Anthropic, Claude Code documentation, "Connect Claude Code to tools via MCP": https://code.claude.com/docs/en/mcp
- Anthropic Help Center, "Get started with custom connectors using remote MCP": https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp
- Anthropic Help Center, "Use connectors to extend Claude's capabilities": https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities
- Anthropic, Claude API documentation, "MCP connector": https://platform.claude.com/docs/en/agents-and-tools/mcp-connector
- Model Context Protocol, "Connect to local MCP servers": https://modelcontextprotocol.io/docs/develop/connect-local-servers
- Model Context Protocol, "Security Best Practices": https://modelcontextprotocol.io/docs/tutorials/security/security_best_practices
Newsletter
One short email when we publish
Notes on AI, image, video and brand work, from the people who do it. No spam.