Claude MCP Setup: Desktop, Web and Claude Code
· DigifyNow
You can connect MCP servers to Claude in four ways. To add a remote server in claude.ai or Claude Desktop, go to "Customize > Connectors" and create a custom connector with a name and URL. The server has to be reachable from the internet. Local servers run on your own computer, either as a .mcpb extension or as an entry under "mcpServers" in claude_desktop_config.json. In Claude Code, you run claude mcp add --transport http with a name and URL. claude mcp list shows what's connected.
What you'll end up with and what you need
By the end of this guide, Claude will use the Model Context Protocol (MCP) to call tools and data from your other systems, such as a ticketing system, a file folder or an ad platform. MCP is an open protocol. An MCP server provides tools, and Claude is the client that uses them. If you want to know how these servers work under the hood, read our article MCP Server: Architecture, Examples and Limits. This guide only covers connecting them. Menu names follow Anthropic's documentation, which we link in each section, and they may change in future versions.
- A Claude account. On the Free plan, you can create one custom connector. On Pro and Max, you set up connectors yourself. On Team and Enterprise, an admin handles it.
- For local servers: Claude Desktop on macOS or Windows. Many servers also need Node.js on the machine.
- For Claude Code: a working Claude Code installation in your terminal.
- The server's address (URL) or launch command, plus login credentials if the server requires authentication.
For small and mid-sized businesses without their own IT team, here's how it works out in practice. A ready-made remote server from a vendor you know takes a few minutes to add. A local server set up through a JSON file takes more time because Node.js, file paths and updates have to be maintained on every single computer. With five workstations, that's five separate installs that can each drift out of sync.
Claude MCP servers: which of the four setup paths fits?
The right path mostly depends on where the server runs. If it's on the internet, Claude connects to it from Anthropic's cloud. That means a server reachable only inside your company network won't work as a connector. If it runs locally, Claude Desktop or Claude Code starts it as a program on your computer.
| Path | Where the server runs | Where you add it | Best for |
|---|---|---|---|
| 1. Custom connector | On the internet, publicly reachable | claude.ai or Claude Desktop, Customize > Connectors | Ready-made services with their own server URL, teams that want nothing installed per computer |
| 2. Desktop extension (.mcpb) | Locally on the computer | Claude Desktop, Settings > Extensions | Local servers offered as a package |
| 3. JSON config | Locally on the computer | claude_desktop_config.json | Local servers without a package, your own servers, testing |
| 4. Claude Code | Locally or on the internet | Terminal, claude mcp add | Development, projects with a shared .mcp.json |
Path 1: Add a remote server as a custom connector
This path works the same way in claude.ai in your browser and in Claude Desktop. The steps are in Anthropic's guide to custom connectors using remote MCP.
- Pro or Max: Open "Customize > Connectors".
- Click "+ Add", then "Add custom connector".
- Enter a name you'll recognize in the chat and paste the vendor's full server URL.
- If the server requires authentication, sign in in the window that opens, using the account whose data Claude should see.
- Team: The Owner or Primary Owner creates the connector under "Organization settings > Connectors". On Enterprise, a custom role with admin access to libraries can also do this.
- Team and Enterprise: Each member then clicks "Connect" and signs in with their own account.
To see what this looks like with a specific vendor, check out our guide How to Set Up Meta Ads MCP in Claude and ChatGPT.
Path 2: Install a local server as a desktop extension
A desktop extension is a local MCP server packaged as a file with the .mcpb extension. Claude Desktop installs it along with its runtime, so you don't have to edit any config file. We give the menu names as they appear in the English interface. They're taken from the documentation and may shift slightly between app versions.
- In Claude Desktop, open Settings and go to "Extensions".
- Click "Browse extensions" and find the server you want.
- Click "Install".
- If the extension asks for settings, such as a folder or an API key, enter them in the dialog before you enable it.
The catch: an extension only runs on the computer where you installed it. If three colleagues need the same server, each one has to install it themselves.
Path 3: Add a local server through the JSON config
For servers that don't come as a package, you enter the launch command by hand. The protocol documentation on connecting local servers and Anthropic's help article on local MCP servers in Claude Desktop cover this in more detail. Menu names here are in English as well.
- In Claude Desktop, open Settings, select "Developer" and click "Edit Config".
- This opens claude_desktop_config.json. On macOS it's in ~/Library/Application Support/Claude/, on Windows in %APPDATA%\Claude\.
- Add an object called "mcpServers". Inside it, give each server a name, and under that name add "command" (the program, for example npx) and "args" (a list of arguments, such as the package name and the folder you want to share).
- If the command is npx or node, check that Node.js is installed.
- Save the file.
- Quit Claude Desktop completely, not just the window, and restart it.
Watch your file paths on Windows: in JSON, every backslash has to be doubled, so write C:\\Users\\… instead of C:\Users\…. A single backslash makes the whole file invalid.
Path 4: Add MCP servers in Claude Code with claude mcp add
In Claude Code, you set up servers in the terminal. All options are listed in the Claude Code MCP reference.
- Add a remote server: claude mcp add --transport http NAME URL, where NAME is any name you choose, such as "tickets", and URL is the vendor's server address.
- Pick a scope: if you don't specify one, the default is local, so the server is active only for you and only in this project.
- With --scope project, the entry goes into a .mcp.json file in the project folder. That file gets checked in, so everyone on the team gets the server.
- With --scope user, the server is available to you in all your projects.
- Start Claude Code and type /mcp. You'll see the status of every server and can start the sign-in if needed.
- To remove a server, run claude mcp remove NAME, using the name you gave it when you added it.
How to check that it's working
In claude.ai and Claude Desktop, a working connector shows up in the list under "Customize > Connectors". To check local servers in Claude Desktop, open the "Add files, connectors, and more /" menu, then "Connectors" and "Manage connectors". The server is listed under the name you gave it in "mcpServers", and you can switch its tools on and off one by one. The real test happens in the chat: ask a question that can only be answered using the server. Before calling a tool, Claude normally asks for your permission. If that prompt appears with the tool's name, the connection works.
claude mcp list: check your servers in Claude Code
In the terminal, claude mcp list shows every server configured for the current project, along with its connection status. If a server is listed but not connected, open /mcp in Claude Code. There you can see whether it's just missing authentication and start the sign-in right away. If a server doesn't appear in the list at all, you probably added it in a different project folder or a different scope.
Common Claude MCP setup errors
| What you see | Cause | Fix |
|---|---|---|
| The local server doesn't show up after saving | Claude Desktop is still running in the background (menu bar or system tray) and hasn't reloaded the file | Quit the app completely and restart it |
| Not a single local server loads | Syntax error in claude_desktop_config.json, such as a trailing comma after the last entry or a single backslash | Validate the JSON and read the error in the mcp*.log files (on macOS in ~/Library/Logs/Claude) |
| The log says npx or node can't be found | Node.js is missing or not on Claude Desktop's search path | Install Node.js, restart Claude Desktop, and if needed enter the full path as the command |
| The connector won't connect | The server is only reachable on the company network or over VPN, but Claude connects from Anthropic's cloud | Make the server publicly reachable with authentication, or use Path 3 or 4 locally |
| Your colleague can't see your server in Claude Code | The server was added in the default local scope | Add it again with --scope project, as long as it contains no credentials |
| The server is listed, but tool calls fail | Authentication expired or was never completed | Open /mcp and start the sign-in again |
Limits and security: what access you're granting
An MCP server acts with the permissions of the account you connect it with. If that account can delete customers, so can a tool call. So connect an account with restricted permissions wherever possible, and read the permission prompt before you approve it. Only trust servers whose source you know. A local server runs as a program on your computer with your user permissions.
- Prompt injection: if a server pulls in outside content, such as web pages, emails or customer tickets, that content can contain hidden instructions aimed at Claude. Don't combine servers like these, unchecked, with servers that can write or send data.
- Credentials: tokens and API keys never belong in .mcp.json or in a repository. According to the documentation, servers with credentials should be added in the local scope.
- No processes without a human in the loop: MCP makes tools available, but it doesn't replace a planned, monitored process. For tasks that should run every night unattended, a chat client is the wrong choice.
- Not a substitute for access control: who gets to see which data is managed in the connected system, not in Claude.
If several employees need access to the same systems, it pays to plan this centrally instead of computer by computer. That's part of what we offer: we help you build a company-wide AI assistant connected to your systems, including training for your team.
Frequently asked questions
What is MCP, in simple terms?
MCP (Model Context Protocol) is an open standard that lets an AI program like Claude call tools in other systems. Think of it as a universal plug: a system's vendor builds an MCP server once, and any MCP-capable client can use it, with no need for a separate integration for every AI.
How much does an MCP server cost?
The protocol itself is open and free. Costs come from your Claude plan, any license for the connected system, and, for your own servers, development and hosting. The cost drivers are the number of tools, the interfaces involved, authentication and hosting. You can only get a reliable figure once the specific scope is clear.
What is the .md file in Claude?
Usually this means CLAUDE.md, a Markdown text file that Claude Code reads when it starts a project. It holds persistent instructions, such as commands, conventions or notes about the project. It has nothing to do with MCP: CLAUDE.md provides knowledge, an MCP server provides tools.
Can I use Claude MCP on the free plan?
Yes, with limits. According to Anthropic's help center (as of October 2026), Free users can create one custom connector. Local servers in Claude Desktop and Claude Code are set up through the config regardless of plan. If you need several remote servers or want to manage them centrally for a team, you need Pro, Max, Team or Enterprise.