Claude MCP Setup Guide

Claude MCP Setup Guide

This guide walks through connecting Claude to an MCP server. Use the local Claude Desktop configuration flow when the server runs on your machine, and use Claude's connector flow when the MCP server is remote and reachable over HTTPS.

The main decision is the transport:

Prerequisites

• Claude Desktop installed and updated if you are configuring a local server.
• A working MCP server, either a local command or a remote HTTPS endpoint.
• Any required API keys or OAuth access configured outside the chat prompt.
• A low-risk read-only tool you can use for the first verification call.

Step 1: Choose Local or Remote

Use a local stdio server when Claude needs controlled access to local files or developer tools. Claude Desktop can start these servers from a JSON configuration file.

Use a remote MCP server when the server is hosted on the internet, shared by a team, or connected to a SaaS application. Remote servers should use HTTPS and should keep authentication on the server or connector side.

Step 2: Configure a Local Claude Desktop Server

Claude Desktop's local configuration file is located at:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

A local stdio configuration uses a command and arguments:

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

Only grant local filesystem access to directories you are comfortable with Claude reading or modifying.

Step 3: Configure a Remote MCP Server

For a hosted API-backed MCP server, add the remote server URL through Claude's connector settings. The URL should be the MCP endpoint provided by the server owner, not a generic website URL or a REST API base URL.

Remote servers commonly require OAuth, API keys, or another authentication flow. Complete authentication through the connector flow instead of pasting secrets into chat.

Step 4: Restart or Reconnect Claude

After changing a local config file, fully quit and restart Claude Desktop so it can load the server. For remote connectors, finish the connector setup flow and confirm the server appears in Claude's connector list.

Step 5: Verify the Connection

Start with a low-risk request:

"What tools do you have available?"

Then run one read-only tool call. If tools do not appear, check:

• The JSON config is valid.
• The command exists on your system path.
• All local paths are absolute.
• Required environment variables are available to Claude Desktop.
• The remote URL is the MCP endpoint and uses HTTPS.
• Authentication has completed successfully.

Common Issues

Security Checklist

• Expose the smallest set of tools Claude needs.
• Prefer read-only tools for the first connection test.
• Keep API keys out of prompts and client-visible config when possible.
• Require approval for write, delete, billing, or admin operations.
• Review MCP logs after the first successful tool call.

Official Sources

• Connect to local MCP servers - official Claude Desktop local configuration and troubleshooting workflow.
• Connect to remote MCP servers - official remote connector flow for Claude and other MCP-compatible clients.
• Claude Code MCP reference - official Claude Code commands, transports, scopes, and MCP troubleshooting behavior.
• MCP transports specification - official stdio and Streamable HTTP transport requirements.

Next Steps

After Claude can see the server, review each exposed tool before using it with production data. For production deployments, combine tool scoping, credential isolation, approval workflows, and audit logging before enabling destructive operations.