MCP Servers on Mac

Written & tested for accuracy by a developer (not AI) · Updated October 2026

MCP Servers on Mac

MCP servers explained: what an MCP server is, what the Model Context Protocol means, and how to use MCP servers safely with Claude Code and Codex on a Mac.

MCP servers give an AI agent abilities it doesn't have on its own. Claude Code and Codex are "AI harnesses" that connect your Mac to an LLM (large language model) and carry out requests to read files and run commands. MCP servers add capabilities to the harness from specialized software and external services, to overcome the limits of the LLM. Examples are fetching web pages (which the LLM can't do very well) or working with specialized software for messaging or productivity.

Here I explain what MCP servers are, how they work, how to use them safely, and how to add them to the AI tools on your Mac.

Connecting MCP servers is one part of getting a Mac ready for coding with AI. See the full Mac setup roadmap for everything else.

Before you get started

I recommend using our free app to set up Claude Code or Codex with MCP servers. It is a FREE Mac app that installs the developer tools and AI coding agents you need, adds MCP servers to Claude Code and Codex, and verifies your setup is working. It does the setup work this guide describes, but easier and faster. Here are all the details about the app before you download. Then download it:

What MCP means

MCP stands for Model Context Protocol. It is an open standard for connecting AI applications to outside tools and data, so any AI app can use any MCP server without custom code. Anthropic, the company that makes Claude, introduced MCP in November 2024. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, part of the Linux Foundation, so no single company controls it. Most AI coding tools now support it, including Claude Code, Codex, Cursor and VS Code.

What an MCP server does

An MCP server is a small program that offers an AI agent a set of tools. The word "server" suggests a computer on a rack in a data center, but many MCP servers are small programs that run on your own Mac. Others run on a company's computers, and your AI app reaches them over the internet.

An MCP server can give your agent three kinds of things:

  • Tools - actions the agent can take, such as searching the web, fetching a web page, or creating an issue in a bug tracker.
  • Resources - information the agent can read, such as files, database records, or documents.
  • Prompts - ready-made instructions that you can run by name.

Tools are the part you'll use most. Here are the kinds of MCP servers a Mac developer is likely to add:

  • Fetch web pages - an AI agent that fetches a web page on its own often gets raw page code, a blocked page, or nothing at all. A web-reading MCP server returns clean text the agent can use.
  • Look up current documentation - AI models learn from data that is months or years old, so they make mistakes with new versions of libraries. A documentation server looks up the current version.
  • Test a web app in a browser - a browser server lets the agent open your app, click through it, and see what goes wrong.
  • Work with GitHub - read issues, open pull requests, and review code.
  • Connect your team's tools - read and update Slack, Notion, Jira, Linear and similar services, using your own sign-in.

Some of these capabilities are also available through a SKILL.md file or the command line.

Learn how MCP works

The AI app you use is the MCP client. Claude Code, Claude Desktop, Codex and the ChatGPT desktop app are all MCP clients. When the app starts, it connects to each MCP server you have set up, asks what tools the server offers, and gives the AI agent a short description of each tool. When the agent decides a tool will help with your request, the app sends the request to the server and passes the result back to the agent. You'll usually see the app ask your permission before the agent uses a tool for the first time.

MCP servers come in two kinds, local and remote.

A local MCP server runs on your Mac. Your AI app starts it with a command, often npx (which needs Node.js) or uvx (which needs uv, the Python tool). A local server runs with your permissions, so it can read your files and run programs, just as you can. That is what makes local servers powerful, and it is also why you should only run ones you trust.

A remote MCP server runs on an external service's own computers, and your AI app connects to it with a web address. Many remote servers ask you to sign in with your own account, so the agent sees only what you can see. Nothing extra runs on your Mac and there is nothing to keep up to date. I recommend a remote server from the service's own company wherever one exists, because it is the simplest to set up and the easiest to trust.

Compare MCP vs an API

An API (application programming interface) is the way one program asks another service for data or tells it to do something. Every service has its own API with its own rules, and a developer writes code for each one. MCP doesn't replace APIs. An MCP server usually sits on top of a service's API and describes its tools in a standard way, so any AI agent can use them without custom code. Think of the API as the service's own language, and the MCP server as an interpreter that every AI agent understands.

As AI agents get more powerful, they can write programs that run on the command line to connect to servers via an API. An MCP server gives the AI agent a ready-made API connection.

AI apps that work with MCP servers

Several AI apps on your Mac can use MCP servers, and each keeps its own list of available MCP servers:

  • Claude Code - the AI coding agent from Anthropic that runs in your terminal. See Claude Code on Mac.
  • Claude Desktop - Anthropic's desktop app, with chat and a Code tab. Its Code tab uses the same MCP servers as Claude Code. See Claude Code vs Claude Desktop.
  • Codex and the ChatGPT desktop app - OpenAI's coding agent. The Codex command-line tool and the ChatGPT desktop app share one list of MCP servers. See ChatGPT for Mac.
  • Code editors - Cursor, VS Code and Zed each have their own MCP settings.

Adding a server to one app doesn't add it to the others (except where noted above). If you use both Claude Code and Codex, add each MCP server to both.

Add your first MCP server

There are two ways to add MCP servers.

The easy way is to let my "Set Up Your Mac" app do it. I built the apps to offer a web-reading MCP server to everyone, and team tools to people who work on a team. The app adds the servers you choose to both Claude Code and Codex, so they work in every project. See Developer Setup: MCP Servers for what the app adds and how to check it, and "Set Up Your Mac" for everything else the apps set up.

You can also add a server yourself with a terminal command. You'll need the server's name and its web address (or, for a local server, the command that starts it) from the server's own instructions.

In Claude Code, the claude mcp add command adds a server. For a remote server, give it a short name and the server's web address. Add --scope user so the server works in every project, not just the folder you're in:

$ claude mcp add --scope user --transport http <name> <url>

In Codex, the codex mcp add command does the same:

$ codex mcp add <name> --url <url>

Restart Claude Code or Codex to load the new server. Then check that it connected. In Claude Code, run claude mcp list and you'll see "Connected", "Needs authentication" or "Failed to connect" beside each server. If a server needs you to sign in, type /mcp inside Claude Code and choose the server. For every other option, including local servers and Claude Desktop, see Add MCP Servers to Claude Code and Claude Desktop.

Find more MCP servers

You'll find thousands of MCP servers listed in public directories, and most of them are written by individuals, not by the company whose service they connect to. Start with the catalogs built into your AI app, which list servers mostly run by the services' own companies:

  • Claude - Anthropic's connectors directory lists remote servers you can connect from Claude's settings. In Claude Code, type /plugin and choose Discover to browse plugins, many of which include MCP servers.
  • Codex - type /plugins in Codex to browse OpenAI's plugin catalog.

For everything else, see MCP Server Directories and Lists, which explains where to look and how to judge a server. For the servers I recommend, see Best MCP Servers for Coding on Mac.

Use MCP servers safely

An MCP server is software you invite into your AI agent, so treat it with the same care as any program you install. There are two risks.

First, a local MCP server is a program running on your Mac with your permissions. A malicious one can read your files or run commands. The MCP project's security best practices warn about exactly this.

Second, anything a server returns becomes part of what your agent reads. A web page, email or document can contain hidden instructions written to trick the agent into doing something you didn't ask for. This is called prompt injection. Anthropic warns that a malicious MCP server may include hidden instructions of its own.

Follow these rules to stay safe:

  • Prefer servers published by the company that runs the service.
  • Prefer remote servers that use your own sign-in over local ones.
  • Give each server the fewest permissions it needs, and start with read-only access where the server offers it.
  • Read what the agent is about to do before you approve a tool, and use "always allow" sparingly.
  • Keep API keys out of Git and out of screenshots.
  • Remove servers you no longer use.

Compare MCP servers with skills and command-line tools

MCP servers aren't the only way to give an agent new abilities. A skill is a folder of instructions (and sometimes small scripts) that the agent reads when a task calls for it. And an agent that can run terminal commands can use command-line tools directly, such as gh for GitHub or git for version control.

Some developers prefer a command-line tool whenever a good one exists, because every MCP server's tool descriptions take up room in the agent's working memory (its "context"). Claude Code now reduces that cost by loading full tool descriptions only when the agent needs them. I suggest using an MCP server when a service needs your sign-in, when no command-line tool exists, or when you work in a desktop chat app that can't run commands. See MCP or Skills or CLI for the full comparison.

Read the MCP guides

Continue setting up your Mac

Don't miss the full visual roadmap and checklist that shows how to set up a Mac for software development, with all the essential tools and settings you might not yet know about.