MCP + A2A + Selenium with Live Demo

[vishalmysore]

Pre-submission Checklist

• This post follows the MCP community guidelines

What would you like to share?

I wanted to share a small project I built while exploring ways to make Selenium automation more approachable. The main challenge was getting A2A (Agent-to-Agent) protocol and MCP (Model Context Protocol) to work together in Java - something I hadn't seen done before.

The Challenge

The problem was pretty straightforward: I had experience with Selenium but was curious if I could:

1. Use natural language to control web automation
2. Make it work with both A2A and MCP protocols
3. Keep it simple enough that others could easily build on it

🎮 Try the Live Demo

Want to see it in action before diving in? Here are some things to try (remember, results may vary based on your prompts):

• Run automated web tests (start with simple navigation tasks)
• Capture screenshots (works best with stable pages)
• Execute natural language commands (be specific and clear in your instructions)
• Watch agents communicate in real-time

💡 Pro Tip: When using the demo, try to:

• Be specific in your instructions
• Start with simple commands and gradually increase complexity
• If something doesn't work, try rephrasing your prompt
• Share what prompts worked best for you!

🚀 What I've Built (And You Can Too!)

I've created a web automation agent that's ready for you to use, extend, and improve. Here's what it can do:

✨ Core Features (Try them live!):

• Executes Selenium-based web automation tasks with natural language
• Captures and validates UI elements automatically
• Communicates between agents using A2A and MCP protocols
• Integrates with your favorite AI models (Gemini, OpenAI, Claude, Grok)
• Provides real-time agent communication and task monitoring

🎯 Live Demo Highlights:

• Test web automation scenarios instantly
• Watch agents collaborate in real-time
• Experiment with different AI models
• No setup required - just visit the demo URL!

How It Works

The core idea is simple: use A2A protocol for handling the automation commands and MCP for AI model communication. Here's the basic flow:

@Action(description = "perform actions on the web with selenium and return text")
public String browseWebAndReturnText(String webBrowsingSteps) throws IOException {
    CustomScriptResult result = new CustomScriptResult();
    A2ASeleniumCallBack seleniumCallBack = new A2ASeleniumCallBack(result);
    // Convert natural language to Selenium actions using AI
    StringBuffer steps = processor.query(webBrowsingSteps);
    script.process(steps, seleniumCallBack);
    return result.getLastData();
}

Main Challenges Solved

1. Protocol Bridge: Got A2A and MCP working together smoothly in Java
2. Error Handling: Built basic recovery for common Selenium issues
3. Simple API: Kept the interface straightforward for easy integration

Basic Setup

If you want to try it out, the setup is minimal:

{
    "webbrowsingagent": {
        "command": "java",
        "args": [
            "-jar",
            "/work/a2a-mcp-bridge/target/mcp-connector-full.jar",
            "http://localhost:7860/"
        ],
        "timeout": 30000
    }
}

Key Components

• Selenium WebDriver: Handles web automation
• A2A Protocol: Manages command structure
• MCP: Connects to AI models
• Spring Boot: Runs the whole thing

You can try the live demo to see how it works.

@Override
public void afterWebAction(String lineProcessed, WebDriver driver) {
    try {
        validateActionResult(driver);
    } catch (StaleElementException e) {
        // Retry with updated element reference
        retryWithRefresh(lineProcessed, driver);
    } catch (TimeoutException e) {
        // Attempt alternative timing strategy
        retryWithExtendedWait(lineProcessed, driver);
    }
}

What I Learned

1. Protocol Integration: A2A and MCP can work together nicely in Java
2. Error Handling: Keep it simple but robust
3. AI Integration: Different models need different prompt structures

The code is on GitHub if you want to take a look or build something similar. I'd be curious to see what others do with this approach.

Setting Up Your Own Instance

If you want to experiment with the code, here's a basic setup to get you started. I've tried to make it as straightforward as possible:

{
    "webbrowsingagent": {
        "command": "java",
        "args": [
            "-jar",
            "/work/a2a-mcp-bridge/target/mcp-connector-full.jar",
            "http://localhost:7860/"
        ],
        "timeout": 30000
    }
}

Note: You might need to adjust timeouts based on your network and the websites you're automating.

2. Remote Server Connection

{
    "webbrowsingagent": {
        "command": "java",
        "args": [
            "-jar",
            "/work/a2a-mcp-bridge/target/mcp-connector-full.jar",
            "https://vishalmysore-a2amcpselenium.hf.space"
        ],
        "timeout": 30000
    }
}

Relevant Links

• GitHub Repository
• a2ajava Library
• MCP Connector Guide
• Demo Server