Install an MCP server: Claude, ChatGPT, or Cursor

To install an MCP server in Claude, ChatGPT, or Cursor, add either a remote server address or the configuration for a program that runs on your computer. The steps depend on the client: check the connection, review the available tools, and perform a read operation before allowing changes.

Prepare the connection and choose your setup

The MCP protocol lets an assistant discover tools made available by a server. You configure that connection in the application; typing an instruction into the conversation is not enough.

Before you start, gather:

an account with permission to add integrations;
the official documentation for your chosen server;
its remote MCP address or local launch command;
the required credentials and a test account;
some nonsensitive data to check that everything works.

To choose a starting point, browse the MCP server listings and the guide to choosing an MCP server. Then return to the provider’s documentation: a server overview does not replace its installation instructions.

Criterion	Local server	Remote server
Where it runs	A program on your computer	A program hosted elsewhere
Configuration	Command, arguments, environment	MCP address and authentication
Requirements	The program’s dependencies	Network access from the client
Access to local files	Depends on the permissions granted	Requires a suitable access mechanism
Maintenance	Your responsibility	Depends on the provider or hosting service

The instructions below reflect the documentation consulted on October 24, 2026. Each setup is separate; do not assume that the same configuration will work in every application. To compare the assistants themselves, see the ChatGPT and Claude comparison.

Install a local server in Claude Desktop

For this example, use the Filesystem server, which provides file operations. Its official repository explains that it can read, write, move, and search files. Giving it access to a folder therefore grants more than read-only access.

Prepare a test folder. Put a few disposable files in it. For this initial setup, avoid your entire home folder, backups, and work documents.
Check the runtime requirement. The official tutorial uses Node.js and provides this command:

node --version

Open the local configuration. In Claude Desktop settings, go to “Developer,” then “Edit Config.” The file you will edit is named claude_desktop_config.json.
Add the configuration. The macOS example below follows the structure in the official local connection tutorial. Replace the paths with the path to your test folder. If the file already contains other servers, keep their entries.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/username/Desktop",
        "/Users/username/Downloads"
      ]
    }
  }
}

Save, fully quit the application, and reopen it. Then review the server’s tools in the connector management interface.

The paths in this example are specific to macOS. On Windows, use appropriate paths and double the backslashes inside JSON strings. Do not copy the suggested folders without first checking what they contain.

Before allowing any write operation, ask the assistant to call list_allowed_directories. This tool lets you check which folders are actually authorized. The server can also receive folders through the MCP Roots mechanism, so verify the effective permissions after connecting.

Add a remote server in Claude

For a hosted service, use the remote connector setup. You do not need to run the server’s program on your computer.

This distinction matters even in the desktop application. According to Claude’s custom connector documentation, the remote connection originates from Anthropic’s infrastructure. An address that is accessible only from your home network is not sufficient.

For an individual Pro or Max account, the documented steps are:

Open “Customize,” then “Connectors.”
Select “+ Add,” then “Add custom connector.”
Give the connector a clear name and enter the MCP address supplied by the provider.
Review the detected authentication method and configure it according to the service’s instructions.
Finish by selecting “Add,” then enable the connector in the conversation.

In a Team or Enterprise workspace, adding a connector for the organization depends on your account permissions. A member may need to wait for an administrator to make the connector available before connecting their own account.

Check the domain before signing in. A repository address, home page, or documentation URL is not necessarily an MCP address. Copy the connection endpoint specified by the provider rather than guessing its path.

For this first test, choose a limited operation, such as finding a test document or looking up a known piece of data. You can then review the server’s create, send, and delete actions separately.

Connect an MCP server to ChatGPT

OpenAI’s current procedure describes adding a server through “Plugins,” without enabling developer mode first. Older tutorials that use “Apps” refer to an earlier interface, which is still documented as a legacy setup.

As of October 24, 2026, that page describes full MCP support for Business and Enterprise/Edu workspaces on the web. Permission to create connections depends on your role, so check your access before looking for a button that may not be available.

Open “Plugins.”
Select the plus button, then “Add custom MCP server.”
Enter a name and, if useful, a description.
Enter the public HTTPS MCP endpoint, or choose “Tunnel” if a compatible tunnel is available.
Configure authentication, review the warning, then select “Create as a plugin.”
Review the discovered tools, install the resulting plugin, and select it with an @ mention in a new conversation.

Here, “install” means adding the connection to ChatGPT. It does not mean that ChatGPT directly runs a local command like the one in the previous example.

Start with a test account that has limited permissions. If you later plan to share the installation with an organization, treat publishing it as a separate step. A working connection does not establish that it is suitable for every user.

Configure a server in Cursor

For a remote example, Context7 provides access to technical documentation. Its official repository lists the MCP address and describes using an API key in an authorization header.

Cursor’s MCP documentation distinguishes between project-specific configuration in .cursor/mcp.json and personal global configuration in ~/.cursor/mcp.json.

Choose the scope you need. A server used by a single repository can stay in that project’s configuration.
Create or open the corresponding file.
Add this remote configuration, which combines Cursor’s documented format with the endpoint published by Context7:

{
  "mcpServers": {
    "context7": {
      "url": "https://mcp.context7.com/mcp",
      "headers": {
        "Authorization": "Bearer ${env:CONTEXT7_API_KEY}"
      }
    }
  }
}

Set your key in the environment available to Cursor, using the variable name shown in the example.
Open “Customize,” then MCP management, and check that the tools appear.

The environment variable keeps the key out of the project file. However, it must be available to the application: a variable set in a separate terminal may not be passed to an editor that is already running.

Next, request a specific documentation lookup. The expected result is a call to the server and usable content, before you ask for any code changes. If you want to develop your own integration, continue with the tutorial on creating your own MCP server.

Check actions and limit permissions

A connection indicator does not validate the entire installation. Check three things separately: the server responds, the expected tool is called, and its result matches data you can verify.

Prepare a brief acceptance checklist recording the server’s name, the source of its configuration, the connected account, the permissions granted, and the expected result. This record can help explain a change in behavior after an update or a new sign-in.

Do not start with a broad request such as organizing all your files or processing every message. Begin with a specific lookup, followed by a reversible write operation in a test area. Compare the result with the actual source rather than relying only on the assistant’s summary.

A retrieved document can also contain a prompt injection: instructions intended to redirect the assistant’s behavior. Content returned by a tool must remain data for the assistant to analyze.

The guide to securing an AI agent explains the permissions and approval checks to consider. For your first installation, focus on defining what the connected account can do and how to remove its access.

Common problems

The local server will not start. Check the client’s logs. Verify the program name, required dependencies, and paths. Rephrasing your request will not fix a launch error.

The configuration file is rejected. Look for a trailing comma, incorrect quotation marks, or an improperly escaped Windows path. Keep a copy of the previous configuration before each change.

The remote server cannot be reached. Check where the connection originates. An address that works in your browser is not necessarily accessible from the assistant’s service.

Authentication succeeds, but data is missing. Check the selected account, workspace, and permissions. Test a resource that the account explicitly has access to.

The assistant answers without calling the server. Request a specific operation and inspect the available tool-call traces. A plausible answer does not prove that the connection was used.

You want to remove the integration. Delete its configuration or connection in the client. If you no longer plan to use it, also revoke the dedicated credentials with the provider.

Before considering the installation complete:

The configuration comes from official documentation.
The server exposes the expected tools.
The account and authorized folders have been checked.
A read operation has returned a verified result.
Sensitive actions have appropriate controls.
You know how to remove the integration and revoke access.

Frequently asked questions

Can I install an MCP server without knowing how to code?

Yes, if the service offers a remote connection or an extension that works with your assistant. Your main tasks are to find the correct address, understand the requested permissions, and check the result. Manual local configuration requires more care.

Why can't my assistant see my local MCP server?

A cloud service cannot directly reach a program running on your computer. You need a local connection supported by the client, an accessible remote address, or a compatible tunnel. Check this distinction before changing your credentials.

Does a free MCP server mean it is free to use?

Not necessarily. The server, the service it connects to, and the assistant may each have separate charges. Even an open-source program running on your computer can call a paid API. Check the terms for each component before using it regularly.

Should I give the assistant my API key in the conversation?

Enter secrets through the authentication mechanism provided by the client or server. A conversation is a poor place to store credentials. If a key has been exposed, revoke it with the provider and replace it.

Install an MCP server in Claude, ChatGPT, or Cursor