How to Set Up Model Context Protocol (MCP) in Claude

Model Context Protocol, or MCP, is an open standard for connecting an AI application to external data, tools, and workflows. In Claude, an MCP connection may expose read-only resources, searches, calculations, or actions that change another system. The setup is straightforward; the important work is choosing the right connection method and granting only the access the workflow needs.

How MCP Fits Together

An MCP client, such as Claude, connects to an MCP server. The server advertises capabilities such as tools and resources. A local stdio server runs as a process on your computer. A remote server is reached over the network, normally through HTTP. Claude’s connectors use the same MCP foundation but provide account-level discovery, authentication, and governance around remote services.

Path 1: Add a Remote Connector in Claude

Remote custom connectors are available in Claude and Claude Desktop. They are configured through the Claude account, and the connection originates from Anthropic’s cloud infrastructure—not from the laptop’s private network. The server must therefore be reachable from Anthropic’s published IP ranges. A private hostname that works only on your VPN will not work as a remote connector without an approved network design.

For Pro and Max accounts

1. Open Customize > Connectors in Claude.
2. Select +, then Add custom connector.
3. Enter the remote MCP server URL supplied by the provider.
4. If required, configure the provider’s OAuth client details in Advanced settings.
5. Finish setup, authenticate, and enable the connector only in conversations that need it.

For Team and Enterprise

An Owner or Primary Owner first adds the service under Organization settings > Connectors. Members then connect their own identities under Customize > Connectors. That separation matters: organization approval determines which connector is available, while the member’s provider identity determines which data and actions the connector can reach.

Path 2: Install a Local Claude Desktop Extension

Desktop Extensions package a local MCP server and its dependencies as an .mcpb file. Anthropic’s current workflow avoids hand-editing JSON for supported extensions:

1. In Claude Desktop, open Settings > Extensions.
2. Select Browse extensions and choose an Anthropic-reviewed tool, or use Advanced settings > Install Extension for an organization-approved .mcpb file.
3. Configure only the folders, accounts, or API settings the workflow requires.
4. Restart Claude Desktop if the tools do not appear, then verify the extension’s settings and permissions.

Team and Enterprise owners can allow or block public extensions and distribute approved custom extensions. A one-click package improves installation, but it does not remove the need to review what the server can read, write, execute, or send across the network.

Path 3: Add MCP to Claude Code

Claude Code manages MCP connections with the claude mcp command. Use HTTP for a remote service and stdio for a trusted local process.

claude mcp add --transport http company-tools https://mcp.example.com/mcp

# In an interactive Claude Code session, authenticate and inspect status:
/mcp

For a remote server that responds with an authentication challenge, Claude Code can complete OAuth from the interactive /mcp screen. The current documentation recommends HTTP for remote MCP services. Use the exact URL and authentication method published by the server owner.

# Everything after -- is the local server command and its arguments.
claude mcp add --transport stdio local-tools -- /absolute/path/to/trusted-mcp-server --allowed-root /absolute/path/to/project

claude mcp get local-tools
claude mcp list

Use an absolute executable path where practical, pin or approve the server version, and scope filesystem arguments to a project rather than a home directory or disk root. Install the runtime required by that specific server; MCP itself does not require a GPU.

Choose the right scope

• Local: available only to you in the current project; appropriate for private experiments and machine-specific paths.
• Project: stored in the project’s .mcp.json so a reviewed configuration can be shared with the team.
• User: available to you across projects; reserve it for broadly useful, trusted services.

Use --scope project or --scope user only when the broader reach is intentional. Review shared project configuration like executable code because accepting a workspace may cause its local server definitions to run.

When Manual Claude Desktop JSON Is Appropriate

Local MCP servers configured in claude_desktop_config.json remain a separate mechanism from remote connectors. Use manual configuration when the local server is not packaged as a Desktop Extension and its publisher documents the command, arguments, and required environment clearly.

{
  "mcpServers": {
    "local-tools": {
      "type": "stdio",
      "command": "/absolute/path/to/trusted-mcp-server",
      "args": ["--allowed-root", "/absolute/path/to/project"]
    }
  }
}

Fully quit and reopen Claude Desktop after a manual configuration change. If the server fails, validate the JSON, executable path, runtime, permissions, and logs. Never paste a real token into a tutorial, screenshot, repository, or support thread.

Security Review Before You Connect

• Verify the publisher and transport. A server can change behavior after installation; monitor updates and remove abandoned integrations.
• Review every requested permission. Limit OAuth scopes, folders, databases, repositories, and write actions to the minimum business need.
• Expect prompt-injection risk. A server that retrieves external content may return malicious instructions. Keep high-impact tools disabled unless the task requires them.
• Approve actions deliberately. Review tool names, inputs, targets, and destructive effects before allowing a call. Use “always allow” only for tools safe to run unattended.
• Separate identities. Use a dedicated service identity for business automation when it improves least privilege, attribution, and revocation.
• Log and reassess. Record connector ownership, scopes, data classes, approval behavior, and an expiration or review date.

Troubleshooting Checklist

For a personal workflow, the goal is a useful connection with narrow access. For an organization, the same setup becomes an identity, information governance, network, logging, and change-management decision. ITECS AI consulting and managed intelligence services help businesses build those integrations with durable controls.

Sources

• Model Context Protocol — Introduction
• Claude Help Center — Custom Connectors Using Remote MCP
• Claude Help Center — Local MCP Servers and Desktop Extensions
• Claude Code Documentation — Connect to Tools with MCP