How to Install an MCP Server: Claude, Claude Code and More

Once you understand what MCP is, the next step is connecting a server so your assistant can use your tools. This guide covers the general approach, then the exact steps for Claude and Claude Code, which have the most complete MCP support, and pointers for other apps.

Step 1: choose a trustworthy server

An MCP server acts with the permissions you give it, so choose as carefully as you would choose software to install:

• Official servers first. Many services publish their own: check the service’s documentation for “MCP”.
• The official MCP Registry at registry.modelcontextprotocol.io lists publicly available servers.
• Your app’s directory. Claude, for example, has a directory of reviewed connectors.
• Open-source projects: look for recent activity, clear documentation and a reputable maintainer.

Read what the server can do. A server that only reads is far less risky than one that can send, delete or pay.

Step 2: remote or local?

• Remote servers run on the provider’s infrastructure. You connect with a URL and usually sign in with OAuth. They are easiest to set up and work across devices.
• Local servers run on your computer, started by the AI app. Use them when the tool must reach local files, a local database or a program on your machine. They often require a runtime such as Node.js or Python.

When both exist, prefer the remote one unless you need local access.

Step 3a: add a connector in Claude

In the Claude apps (web, desktop and mobile), MCP servers appear as connectors. There are three ways to add one:

1. From the directory: browse the reviewed connectors and add the one you want.
2. As a custom connector: paste the URL of a remote MCP server. Anthropic does not review custom connectors, so only add servers you trust. Which plans can add custom connectors depends on your Claude plan and, on Team and Enterprise, on your organisation’s settings.
3. Inside a plugin: some plugins bundle a connector together with skills that teach Claude how to use it.

You can manage your connectors at claude.ai/customize/connectors. For local servers, the Claude desktop app uses desktop extensions (packaged with the MCPB format), which install like an app instead of requiring you to edit configuration files.

Step 3b: install a server in Claude Code

Claude Code, Anthropic’s coding agent, manages servers from the command line.

Remote server (recommended when available):

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
  --header "Authorization: Bearer YOUR_GITHUB_TOKEN"

GitHub’s remote server signs in with a personal access token: create a fine-grained one with access only to the repositories you need. If a server uses OAuth instead, such as Sentry’s, run /mcp inside Claude Code (or claude mcp login sentry) to sign in through your browser.

Local server (note the -- separating Claude’s options from the server’s own command):

claude mcp add --transport stdio --env AIRTABLE_API_KEY=YOUR_KEY airtable -- npx -y airtable-mcp-server

Choose a scope:

Scope	Available in	Stored in
local (default)	Only the current project, only you	~/.claude.json
project	The current project, shared with your team via version control	.mcp.json in the project
user	All your projects	~/.claude.json

claude mcp add --transport http sentry --scope project https://mcp.sentry.dev/mcp

Manage servers:

claude mcp list          # all servers and their status
claude mcp get github    # details of one server
claude mcp remove github # remove it

A project-scoped .mcp.json file looks like this, and can read secrets from environment variables so you never commit them:

{
  "mcpServers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/",
      "headers": { "Authorization": "Bearer ${GITHUB_TOKEN}" }
    }
  }
}

If you already configured servers in the Claude desktop app, claude mcp add-from-claude-desktop imports them (macOS and WSL). Connectors you added in claude.ai are also available in Claude Code when you are signed in with the same account.

Step 3c: other apps

Most AI apps that support MCP follow the same idea — a server name plus either a URL (remote) or a command (local) — but the settings screen and file format differ:

• ChatGPT and the OpenAI API support remote MCP servers; see OpenAI’s MCP documentation.
• Visual Studio Code and Cursor support MCP in their AI features, configured through their settings or an mcp.json file.

Check each app’s current documentation, as menus change often.

Step 4: authenticate with least privilege

When asked to sign in or provide a token, grant only what the task needs: read-only access, a single repository, a specific workspace. Never paste tokens into chats, and keep them out of files you share or commit.

Step 5: test safely

Start with a harmless request: “List my open GitHub issues” or “What tables are in this database?”. Confirm that the assistant uses the right server and that results look correct before asking it to create, update or send anything.

Step 6: keep it under control

• Leave approval prompts on for actions that write, send or delete.
• Be cautious with servers that read outside content such as web pages or emails, which can contain prompt injection.
• Review your servers periodically and remove the ones you no longer use.
• In Claude Code, project servers from a shared .mcp.json require your approval before they run — read that file before trusting it.