Guides / AI basics

What are MCP servers, and why does your GTM stack need them?

MCP is the standard way to plug an AI assistant into the tools your team already uses. Here is what it is, how it works and how to add your first server.

Nicolas de la Guardia Nicolas de la Guardia · Co-founder · Engineering & data10 min read · checked

An MCP server is a small connector that lets an AI assistant like Claude read from and act in another tool, such as your CRM, a database or Figma. MCP stands for Model Context Protocol, an open standard, so one connector works with any assistant that speaks the protocol instead of each tool needing a custom integration for each AI.

What you need

  • Claude Code installed, or the Claude desktop or web app on a paid plan
  • An account in the tool you want to connect
  • Ten minutes, and permission to connect that tool to an AI assistant

The problem MCP solves

On its own, an AI assistant only knows what you paste into the chat. Ask it to check which open deals have no next step and it cannot, because it cannot see your CRM. Teams work around this by exporting CSVs, copying records into prompts and pasting the answers back. That works once. It does not work as a daily operating habit, and every copy-paste is a chance to leak data or lose context.

MCP replaces the copy-paste with a direct, permissioned connection. The assistant asks the MCP server for what it needs, the server calls the tool's API with your credentials, and only the result comes back into the conversation.

How an MCP server works

There are three pieces:

  • The client is the assistant you talk to: Claude Code, the Claude app or another MCP-capable tool.
  • The server is the connector. It describes a list of tools the assistant may call, such as "search contacts", "get deal" or "get design context".
  • The tool behind it is the real system: HubSpot, Slack, Figma, a Postgres database or your own internal API.

When you ask a question, Claude looks at the tools the server offers, decides which to call, and asks your permission before it runs anything you have not already approved. Servers come in two shapes. A remote server lives at a URL and you sign in with OAuth, so there is nothing to install. A local server is a small program that runs on your machine and usually reads an API key from an environment variable.

What GTM teams connect first

The useful connections follow the work, not the tool list. The ones we see pay off fastest:

  • CRM for pipeline reviews, stale-deal checks and data hygiene. See CRM hygiene with Claude Code.
  • Slack so Claude can read a thread for context or post a summary where the team already works.
  • Enrichment and data tools so research and list building happen in one place instead of five tabs.
  • Figma for landing pages, one-pagers and campaign assets. See Connect Figma to Claude.
  • Call recordings so discovery notes and objections feed back into messaging.

Connecting tools is the first half. The second half is giving Claude the context to use them well, which is what a CLAUDE.md setup for GTM and a shared GTM brain are for.

Add your first MCP server to Claude Code

  1. Find the server's address or package. Most vendors now publish an official MCP server in their developer docs. A remote server gives you a URL. A local one gives you a command, often an npx package.
  2. Add it with one command. For a remote server, pass the URL with the HTTP transport. For a local server, pass the command after -- and any API key with -e. 
    # a remote server (you sign in with OAuth afterwards)
claude mcp add --transport http figma https://mcp.figma.com/mcp

# a local server that needs an API key
claude mcp add my-crm -e CRM_API_KEY=your-key -- npx your-crm-mcp-server
  3. Choose who gets it. By default a server is added for you, in the current project only. Add -s user to use it in every project, or -s project to share it with the team through a .mcp.json file in the repository.
  4. Sign in and check it. Start Claude Code, run /mcp to authenticate remote servers, then list what is connected. 
    claude mcp list
claude mcp get figma
  5. Ask for something small. "List the five most recently updated deals" or "summarise the last message in #sales" proves the connection works before you build anything on top of it.

In the Claude desktop or web app the same idea lives under Settings, Connectors: pick a listed connector or add a custom one by pasting the server URL.

Before you connect a live system

Start read-only. Give the first connection a key or a user role that can look but not change, prove it is useful, and only then allow writes, one action at a time.

  • Use official servers where they exist. A community server runs code on your machine with your credentials. Read what it does before you install it.
  • Scope the credentials. Create a dedicated API key or integration user for the assistant so its access can be limited, audited and revoked on its own.
  • Keep approvals on for writes. Let Claude read freely, but keep it asking before it updates records, sends messages or deletes anything.
  • Keep secrets out of the repository. Put keys in environment variables, not in a shared .mcp.json.
  • Watch the rate limits. Many servers inherit the vendor's API limits and some plans cap calls tightly.

Want your stack connected and governed properly from day one? Book a GTM Engine Review and we will map what to connect first.

Frequently asked

Is MCP only for Claude?

No. Anthropic introduced it, but it is an open protocol and many assistants and editors support it, so a server you add once can often be reused across tools.

Do I need to write code to use MCP?

No. Adding an existing server is one command or a settings screen. You only write code if you want a server for an internal system that nobody has built yet.

Is it safe to connect my CRM?

It can be, if you treat the assistant like a new team member: its own scoped credentials, read-only to start, approvals on for changes and a clear record of what it did.