MCP Setup Guide•Installation, tool usage, and safety checklist

How to Install and Use Model Context Protocol (MCP) Tools in Cursor, Claude, GitHub Copilot, Windsurf, and Replit Agent

This guide shows you how to connect MCP servers to your editor or agent host, how MCP tools behave in chat, and the most important safety notes to keep in mind when you enable new integrations.

1. Getting started with MCP in your editor or agent host

MCP is an open protocol that lets you connect AI agents to external tools, APIs, and data sources through MCP servers. The exact setup flow depends on whether you are using Cursor, Claude, GitHub Copilot, Windsurf, or Replit Agent, but the core steps are the same: add a server, configure transport, and (optionally) provide environment variables such as API keys.

Cursor: install an MCP server in Composer

Cursor uses MCP servers to power tools inside Composer, so the agent can call APIs and services on your behalf.

  1. Open Cursor Settings and navigate to the Features → MCP section.
  2. Click “Add New MCP Server”.
  3. Give the server a nickname (for example, "My Analytics MCP").
  4. Select the transport type (stdio or SSE).
  5. For stdio, enter a valid shell command that starts the server. For SSE, enter the full URL of the SSE endpoint.

Claude Desktop: connect MCP servers to conversations

  1. Open Claude Desktop.
  2. Go to Settings (macOS: Claude → Settings, Windows: File → Settings).
  3. Find the “MCP Servers” or “Model Context Protocol” section.
  4. Click “Add MCP Server”.
  5. Provide a name, pick the transport type (stdio, SSE, or HTTP), and enter either an execution command or an endpoint URL.
  6. If the server requires secrets, add environment variables (like API keys or tokens) in the configuration.
  7. Restart Claude Desktop to ensure the new server is fully loaded.

GitHub Copilot: enable MCP tools in your IDE

  1. Open your IDE settings and navigate to GitHub Copilot configuration.
  2. Alternatively, open the Command Palette (Ctrl+Shift+P /Cmd+Shift+P) and search for “GitHub Copilot: Open Settings”.
  3. Locate the MCP / Model Context Protocol section.
  4. Click “Add New MCP Server”.
  5. Provide a name, choose the transport type, and enter either a command (stdio) or URL (HTTP/SSE).
  6. Add any required environment variables in the configuration UI.

Windsurf: wire MCP into the AI assistant

  1. Open Windsurf and go to Settings or use the Command Palette.
  2. Search for “MCP” or “Model Context Protocol”.
  3. Click “Add New MCP Server”.
  4. Set a nickname, choose stdio/SSE/HTTP, and configure the server command or URL.
  5. Add environment variables if your server relies on external APIs or credentials.
  6. Restart Windsurf so the AI assistant can see the new tools.

Replit Agent: expose MCP tools to your workspace

  1. Open your Replit workspace.
  2. Go to Settings or open the Command Palette and search for MCP configuration.
  3. Click “Add New MCP Server”.
  4. Provide a name, pick the transport type, and set either a command or an endpoint URL.
  5. Configure any required environment variables.
  6. Restart Replit Agent to pick up the new server configuration.

2. Using MCP tools during conversations

Once a server is configured, your agent host exposes its tools to the model. In most products, tools are used automatically when helpful, but you can also nudge the model to use a specific MCP tool by name.

Tool availability

After you add a server, it appears in the MCP servers list for your app. You may need to refresh the app or restart it once so that newly added tools are visible to the agent.

Prompting the agent to use a tool

You can often encourage tool use explicitly by referencing the MCP integration:

  • “Use the [MCP name] tool to fetch the latest data.”
  • “Call the MCP integration that syncs issues from GitHub for this repo.”
  • “Query analytics using the [MCP name] server before answering.”

Typical tool execution flow

While UI details vary by product, MCP tools usually follow the same pattern:

  • The agent proposes a tool call with structured arguments.
  • The UI shows a confirmation step or at least a visible tool call in the transcript.
  • The MCP server executes the action or query and returns structured output.
  • The agent incorporates that output into its next response.

3. Important notes and safety checklist

MCP gives agents powerful access to your systems. Treat each server like production infrastructure and apply the same safety and observability standards you would for other internal services.

  • MCP tools may not be supported by every model or every host—check compatibility before relying on them.
  • Many editors only expose MCP tools to the agent-style workflows (for example, Composer) rather than basic chat.
  • For servers that require secrets, prefer wrapper scripts or environment files so API keys are not hard-coded.
  • Start with read-only tools when possible, then add write actions once you have logging and review in place.
  • Log MCP tool calls (with secret redaction) so you can audit behavior and debug misconfigurations.

Next step: browse the MCP directory for servers you can plug into your stack, or read the Model Context Protocol architecture guide to understand how hosts, clients, and servers fit together.