Skip to main content

Building "Manus" Level Intelligent Agents: Using MCP Protocol to Connect Claude Code and Cursor to the Real World

Building "Manus" Level Intelligent Agents: Using MCP Protocol to Connect Claude Code and Cursor to the Real World

Have you ever noticed that your AI assistant seems to be becoming more and more like you? It not only carries your knowledge system but even begins to replicate your thinking preferences. This feeling is like meticulously crafting a parallel version of yourself—one that understands your workflow, anticipates your needs, and operates with your style.

However, this parallel self has a fundamental limitation: it cannot truly interact with the world. It remains trapped within the confines of your local machine, unable to access real-time information, browse the web, or interact with external services.

This is where the Model Context Protocol (MCP) comes in. MCP transforms your AI from a passive code completion tool into an active problem-solving partner capable of reaching beyond the local environment.

1. Why We Need MCP

Consider this scenario: you ask your AI assistant to check the latest news or look up current information. Here's what typically happens:

Help me check for recent news?

Thinking...

The assistant recognizes your request but must respond:

  • Sorry, I cannot help you check for news. I am a software engineering assistant and can only access your local file system, without the ability to browse the internet or get real-time news.

If you need to check for news, I suggest you:

  • Directly visit news websites
  • Use a search engine
  • Use dedicated news applications

This limitation illustrates what we call the "M×N dilemma": every application (M) that wants to connect to different AI systems (N) requires writing specific integration code. This creates a barrier that traps AI locally, making it an information island.

MCP solves this problem by providing a unified, open protocol—like a universal adapter that opens the door for AI to the outside world. With MCP, your AI is no longer a "standalone tool" limited to reading and writing local files.

Instead, it can:

  • Connect to GitHub to create pull requests
  • Query production databases
  • Read Jira tasks
  • Send Slack messages
  • Browse the web and retrieve real-time information

This is not just an addition of capabilities, but a qualitative leap—evolving from a tool into a truly intelligent agent.

2. Understanding MCP Architecture

MCP architecture can be understood in two parts: "Basic Components" and "Operating Architecture."

Think of it this way: Operating Architecture (Host/Client/Server) is the "logistics pipeline," while Basic Components (Tools/Resources/Prompts) are the "goods" transported through that pipeline.

2.1 Basic Components: Equipping Your Agent with Capabilities

No matter how complex an Agent you want to build, you only need to understand three fundamental components: Tools, Resources, and Prompts.

ComponentPrimitiveFunctionIn Simple Terms
Hand 👋ToolsExecutable functionsYour agent's ability to take action. It can create a GitHub PR, execute a database query, or perform any actionable task.
Eye 👀ResourcesRead-only data sourcesYour agent's ability to observe and gather information. It can read Jira task descriptions, view API documentation, or access any read-only data source.
Mind 🧠PromptsReusable interaction templatesYour agent's thinking and decision-making framework. It follows structured processes, such as "first research, then code, finally test."

Tools are AI's "weapons" or "instruments"—controlled by your AI assistant (Claude Code or Cursor), they are essentially a set of executable functions that allow AI to decide when to call them. Both Claude Code and Cursor include built-in Filesystem MCP with a set of tools. In Claude Code, you can view them by typing /mcp. In Cursor, MCP servers are configured through settings. These tools enable your AI assistant to view and modify local files at any time.

Resources are read-only data sources that provide context for your agent. Because data sources can be sensitive, they are generally controlled by applications or users rather than being directly modifiable by the AI.

Prompts are Standard Operating Procedures (SOPs). They represent experience and methodology rather than intelligence itself (which comes from the model). Prompts break down complex tasks and teach AI step-by-step processes: "what to do first, what to do second."

2.2 Operating Architecture

The MCP operating architecture follows a familiar pattern similar to how computers connect to external devices:

  • 💻 MCP Host: Your main system (like Claude Code or Cursor)
  • 🔌 Server: External services (like GitHub, databases, or web browsers)
  • 🔗 Client: The connection layer between Host and Server (like USB cables or network connections)

The MCP Host (such as Claude Code or Cursor) acts as the central hub, connecting to multiple MCP Servers (like GitHub Server, Sentry Server, or Database Server) through Clients. Each Host can have multiple Clients, and each Client connects to one Server in a 1:1 relationship.

MCP Host: Claude Code or Cursor itself
Server: Filesystem, Chrome-devtools, or any other MCP server
Client: The connection layer (when you see "Connected", it means the Client has successfully connected to the Server)

3. Practical Guide: Configuring Your First MCP

Before configuring MCP, identify your specific needs. What capability do you want to add to your AI assistant?

For this guide, we'll focus on a common need: enabling your AI assistant (Claude Code or Cursor) to browse the web and access real-time information.

3.1 Finding the Right MCP

Many developers have released useful, free, open-source MCP servers on GitHub. You can configure them according to your needs.

To enable browser automation and web access, we'll use Chrome-devtools-mcp, which provides comprehensive browser control capabilities.

The Chrome-devtools MCP provides the following tools:

Tools Available:

  • Input automation (8 tools): click, drag, fill, fill_form, handle_dialog, hover, press_key, upload_file
  • Navigation automation (6 tools): close_page, list_pages, navigate_page, new_page, select_page, wait_for
  • Emulation (2 tools): emulate, resize_page
  • Performance (3 tools): performance_analyze_insight, performance_start_trace, performance_stop_trace

These tools enable your AI to fully control a browser, making it capable of navigating websites, filling forms, clicking buttons, and retrieving information—essentially giving it the ability to "surf the internet."

3.2 Installation and Testing

MCP Client Configuration

Installation methods differ slightly between Claude Code and Cursor. Choose the instructions for your platform:

For Claude Code

On the Chrome-devtools-mcp GitHub page, you'll find this command:

claude mcp add chrome-devtools npx chrome-devtools-mcp@latest

Simply run this command in your terminal, and the installation will complete. Then open Claude Code and enter /mcp to verify the installation. You should see:

Manage MCP servers
2 servers

> 1. chrome-devtools ✔connected · Enter to view details
2. filesystem      ✔connected · Enter to view details

If you see both servers connected, congratulations! Your Claude Code now has browser automation capabilities.

For Cursor

Cursor uses a configuration file approach for MCP servers. You'll need to add the Chrome-devtools MCP server to your Cursor MCP configuration:

  1. Locate or create the MCP configuration file:

    • Global Configuration (recommended):
      • Windows: %USERPROFILE%\.cursor\mcp.json
      • macOS/Linux: ~/.cursor/mcp.json
    • Project-Specific Configuration:
      • Create a .cursor directory in your project root (if it doesn't exist)
      • Add mcp.json file in .cursor/mcp.json

  2. Add the Chrome-devtools MCP server configuration:

{
  "mcpServers": {
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest"]
    }
  }
}

  1. Alternative: Use Cursor Settings UI:

    • Open Cursor Settings (Cmd/Ctrl + ,)
    • Search for "MCP" or "Model Context Protocol" in settings
    • If available, use the UI to add the MCP server with:
      • Name: chrome-devtools
      • Command: npx
      • Args: -y chrome-devtools-mcp@latest

  2. Restart Cursor after configuration. The Chrome-devtools MCP server should now be available to your AI assistant.

Note: If you're using a project-specific configuration, make sure the .cursor directory and mcp.json file are properly set up in your project root.

3.3 Testing the Capabilities

Once installed, you can test the new capabilities. Ask your AI assistant (Claude Code or Cursor) to retrieve information from the web:

> Can you check the latest tech news for me?

Your AI assistant will now:

  1. Recognize that it needs to access the internet
  2. Automatically open a browser using the Chrome-devtools MCP
  3. Navigate to a news website
  4. Retrieve and summarize the information for you
.: Thinking...
The user wants to see the latest tech news. I can use browser tools to help access news websites. Let me open a news site to get the latest information.
● I'll help you check the latest tech news. Let me open a news website.
• chrome-devtools - new_page (MCP) (url: "https://techcrunch.com/")
#new_page response
## Pages
0: about:blank
1: https://techcrunch.com/ [selected]
.: Thinking...
Page loaded, let me get a snapshot of the page content to view the news.

The AI can also handle more complex queries, such as searching for specific information:

> Has A16Z recently raised any funding? How much?

When one search method is blocked or unavailable, the AI will intelligently try alternative approaches, switching between different search engines or methods until it finds the information you need.

This demonstrates the power of MCP: your AI assistant becomes proactive, resourceful, and capable of navigating the complexities of the real world.

4. Conclusion: The Era of Accessible "Manus" Level Agents

From the concept of a "universal adapter" to the collaborative work of "hands, eyes, and brain," and then to the powerful effects of "combined use," MCP represents a blueprint for AI deeply integrated into the entire software development process.

MCP is no longer just a protocol—it's a new work paradigm. It transforms AI from a passive "code completion tool" into an active "problem-solving partner." Through MCP, you can gradually fine-tune and expand your local development environment into a powerful, efficient, and highly customized "Manus"-level intelligent agent.

The capabilities demonstrated in this article represent just the beginning. Advanced features such as multi-turn MCP calls within a single task, integration with authentication services, and collaborative techniques with Skills and Sub-agents open up even more possibilities for AI-assisted development.

The future of software development is one where AI agents seamlessly bridge the gap between local development environments and the broader digital world, making every developer more productive and capable.

Comments

Post a Comment

Popular posts from this blog

Leadership VS. Management

Characteristics and Behaviors of a Good Manager A good manager embodies a range of characteristics and behaviors that foster a productive, motivated, and cohesive team. Below is an exploration of key qualities— professionalism , long vision , leadership skill , honesty , empowerment , and continuous learning —detailing their characteristics, behaviors, and importance. 1. Professionalism Professionalism is the foundation of a manager’s credibility and sets the tone for workplace culture. Characteristics : Consistency : Maintains a steady, respectful demeanor, ensuring fairness. Respectful Communication : Uses clear, constructive, and inclusive language. Appearance and Demeanor : Reflects organizational values, setting a positive example. Boundary Setting : Balances approachability with authority. Behaviors : Models punctuality and reliability in deadlines and commitments. Handles conflicts diplomatically, focusing on solutions. Maintains confidentiality to build trust. Adapts communicat...
Post a Comment
Read more

Five Skills Positive Impact on My Life

I am committed to lifelong learning and seek to acquire diverse knowledge and skills to enhance my personal and professional life. Here are five skills that have had a positive impact on my life:  1. Zero Second Thinking - The world's simplest training for improving your mind .  I learn the skill of Zero Second Thinking from the book of Zero Second Thinking by Yuji Akaba . When I need to come up with an idea or solve a difficult problem, I stop what I'm doing and find a quite place to lay down some white paper.  This is how Zero Second Thinking works: On the first paper, write down 3~5 most important and overview ideas in your mind in one minute, no mater if they are mature or not.  At the beginning, the time doesn't need to be measured accurately, just keep writing until you have listed all the overview points. After serval training, you will know your own speed.  In my minute, I can usually list 7 items.  Take one of the ideas you just write down an...
Post a Comment
Read more

Meanings of Report

 Reports are necessary in all companies and organizations. A effective organization should know what is the true meaning of report and try to avoid useless reports.  There are five most important meanings of a report: 1. Managers collect data to know the situation of business running.  2. Recognize data changes in reports. Changes are related with historical strategies, market trending, etc, which help manager to make new decisions.  3. Extraordinary data, which always hide a risk or an opportunity.  4. Problems and asking helps.  Problem is required to be report to manager, meanwhile one or two possible solution should come with problems.  So that manager can not how to help you and your team to resolve the problem but not compliant you as a problem maker.  Some ineffective reports come with evident characters, 1. No audience. Sometimes manager ask a report for curiousness. But after a few times the manager lost interesting on that and ignore dat...
Post a Comment
Read more