How to Integrate MCP Servers with Claude Desktop, Cursor, and Other AI Clients: A Complete Setup Guide

You can integrate MCP servers with Claude Desktop, Cursor, and other AI clients by configuring a mcpServers.json file that points to the server's HTTP endpoint, enabling automatic tool discovery and invocation via JSON-RPC.

The Model Context Protocol (MCP) standardizes how AI clients discover and execute external tools. According to the punkpeye/awesome-mcp-servers repository, integrating MCP servers allows Claude Desktop, Cursor, Windsurf, and VS Code to automatically invoke capabilities like browser automation, screenshots, and database queries directly from chat interfaces.

Understanding MCP Server Architecture

MCP servers expose capabilities as stateless JSON-RPC endpoints. Each server publishes three core methods: list (tool catalog), get (tool schema), and call (execution). This lightweight protocol enables any language to implement a server or client without complex authentication layers.

Most servers run locally on localhost:3000 or localhost:4000 and expose their tool manifests at the /tools/list endpoint. Clients automatically fetch this manifest to understand available capabilities, creating a plug-and-play experience where the same configuration works across Claude Desktop, Cursor, Claude Code, and Windsurf.

Step-by-Step Integration Process

Install and Run the MCP Server

MCP servers typically distribute as single-command packages via npm or pip. The server starts a local HTTP endpoint that publishes its tool list immediately.

# Example: Install and run Pagebolt MCP for screenshots

npx -y pagebolt-mcp

# Alternative: Python-based servers

pip install aimarket-mcp-packager

# Verify the server is running

curl http://localhost:3000/tools/list | jq .

According to README.md line 435 in the repository, the Pagebolt MCP server supports screenshots, PDFs, and OG-image generation and works across Claude Desktop, Cursor, and Windsurf without additional configuration.

Configure the mcpServers.json File

Create a mcpServers.json file containing an array of server objects. Each object requires a url, optional name, description, and a list of enabled tools.

[
  {
    "name": "Pagebolt – Screenshots",
    "url": "http://localhost:3000",
    "description": "Take screenshots, PDFs and OG‑image generation",
    "tools": ["screenshot", "capture_pdf", "og_image"]
  },
  {
    "name": "Auto‑Browser",
    "url": "http://localhost:4000",
    "description": "Playwright‑based browser automation",
    "tools": ["browser_open", "click", "type", "read_page"]
  }
]

As documented in README.md line 39, Claude Desktop specifically looks for this file format (or .mcpb bundles) to populate its MCP Servers settings panel.

Add to Claude Desktop or Cursor

Open Settings → MCP Servers in Claude Desktop and drag-drop your mcpServers.json file into the UI. The client validates the endpoint, fetches the tool manifest from /tools/list, and makes the tools available in any chat session.

For Cursor and other compatible clients (including Claude Code and VS Code extensions), import the same JSON configuration. The repository notes at README.md line 233 that WayStation-ai's connector completes this process in under 90 seconds, automatically handling the endpoint validation and tool registration.

Invoke Tools via Natural Language

Once configured, simply type requests matching tool names in your AI client. The client automatically translates natural language into JSON-RPC calls to the MCP server.

User: Take a screenshot of https://example.com
Claude Desktop → automatically calls the `screenshot` tool on the Pagebolt server
Result: Returns a PNG image URL displayed in chat

This zero-code invocation works because the client maps your request to the tool schema retrieved during the discovery phase.

Aggregate Multiple Servers (Optional)

For managing multiple servers, use awesome-mcp-tools to aggregate dozens of MCP servers into a single endpoint. This simplifies client configuration by exposing one URL instead of many.

# Install the aggregator globally

npm i -g awesome-mcp

# Start the unified endpoint

awesome-mcp start --port 5000

# Now add http://localhost:5000 to your mcpServers.json

Reference README.md line 165 for implementation details on this aggregation pattern.

Key Configuration Files in the Repository

The punkpeye/awesome-mcp-servers repository contains several critical files for integration:

• README.md: Central index of all MCP servers including install commands and integration notes for Claude Desktop and Cursor. Contains the quick-setup video reference at line 39 and Pagebolt examples at line 435.
• README-zh.md: Provides the same integration guidance for non-English speakers, useful for global teams implementing MCP across different regions.
• CONTRIBUTING.md: Describes how to add new MCP server entries, ensuring future contributions maintain accurate integration instructions for AI clients.

Summary

• MCP servers expose tools via stateless JSON-RPC endpoints at /tools/list, enabling automatic discovery by AI clients.
• Configuration requires a mcpServers.json file with server URLs, names, and enabled tool lists, which works across Claude Desktop, Cursor, Windsurf, and VS Code.
• Installation typically uses single commands like npx -y pagebolt-mcp or pip install, starting local servers on ports 3000 or 4000.
• Invocation happens through natural language in chat interfaces, with clients automatically translating requests into JSON-RPC calls.
• Aggregation tools like awesome-mcp allow you to bundle multiple servers into one endpoint for simplified management.

Frequently Asked Questions

What is the Model Context Protocol (MCP)?

The Model Context Protocol is a standardized JSON-RPC interface that allows AI clients to discover and invoke external tools. MCP servers expose methods for listing available tools, retrieving schemas, and executing functions, creating a universal integration layer that works with Claude Desktop, Cursor, and other AI clients without custom code for each tool.

Do MCP servers require authentication?

Many MCP servers operate in zero-auth mode and require no API keys for local-first tools. Public servers like pagebolt-mcp and vibie-mcp allow direct HTTP access, though they may implement rate limits or optional token-based payment (x402). Authentication is not mandatory for basic local tool integration.

Can I use the same MCP configuration across different AI clients?

Yes. The same mcpServers.json configuration file works with Claude Desktop, Cursor, Claude Code, Windsurf, and VS Code extensions because they all implement the same MCP specification. This cross-client compatibility means you configure your tools once and use them across any MCP-aware IDE or chat interface.

How do I troubleshoot connection issues with MCP servers?

First, verify the server is running by calling curl http://localhost:3000/tools/list and checking for a valid JSON response. Ensure your mcpServers.json URL matches the actual port where the server started. If Claude Desktop fails to discover tools, check that the JSON syntax is valid and that the tools array contains strings matching the server's published method names.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}