Build a Web-Browsing AI Agent in 50 Lines Using Tiny Agents and Playwright MCP

In the fast-moving world of AI tools, developers want two things: power and simplicity. That’s exactly what Tiny Agents offer—smart AI agents you can build in just 50 lines of code.

In this guide, we’ll walk through building a web-browsing agent that can open a website and fetch its title, using:

• Tiny Agents (by Hugging Face)
• Playwright MCP (by Microsoft)

What Are Tiny Agents?

Tiny Agents are small AI programs built around:

• a system prompt
• a language model (LLM)
• and one or more tool servers via Model Context Protocol (MCP)

Instead of stuffing everything into a huge framework, you keep it clean and modular. Tools can be anything from a file reader to a browser or a custom API.

Why Use Playwright MCP?

Playwright MCP gives your agent the power to:

• Launch a browser
• Visit any website
• Click, type, or scroll
• Read page content or take screenshots

What You’ll Need

Before writing code, make sure you’ve got:

• Python installed
• Node.js and npm (for Playwright MCP)
• An OpenAI API key (or any LLM with function-calling support)
• Packages installed:

  pip install tinyagent litellm

Build Your First Web Agent

Let’s create a simple agent that navigates to https://www.example.com and tells you the page title.

The Code

from tinyagent import TinyAgent
import asyncio

async def main():
    agent = TinyAgent(model="gpt-4.1-mini", api_key="YOUR_OPENAI_API_KEY")
    await agent.connect_to_server("npx", ["@playwright/mcp@latest"])
    
    task = "Navigate to https://www.example.com and tell me the title of the page."
    result = await agent.run(task)
    
    print("Result:", result)
    await agent.close()

asyncio.run(main())

How It Works

• TinyAgent connects to your model and tool server (Playwright MCP).
• It sends the task to the model: "Navigate to X, tell me Y."
• The model decides what tools to use (e.g., browser_navigate, browser_snapshot).
• The MCP server runs those tools and returns data.
• The result is displayed and the agent shuts down.

How to Run It

Step 1: Start the browser MCP server

npx @playwright/mcp@latest

To run in background/headless mode:

npx @playwright/mcp@latest --headless

Step 2: Run your script

Save your script as web_agent.py and run:

python web_agent.py

Expected Output:
Result: The title of the page is "Example Domain".

Go Further: Advanced Use Cases

• Search & Summarize: Ask it to search Google and summarize results.
• Form Fillers & Automation: Log in, submit forms, navigate flows.
• Vision Mode: Enable with

  npx @playwright/mcp@latest --vision

  for screenshot-based tasks.

Benefits of Tiny Agents

• Simple: Minimal setup, no boilerplate
• Modular: Swap tools and models easily
• Great for prototyping: Perfect for quick experiments

Limitations to Keep in Mind

• Not ideal for large, multi-step workflows
• Depends on external API (OpenAI, etc.)
• Snapshot mode may not work on very dynamic websites

Troubleshooting Tips

• MCP errors? Start the server before running the script.
• API errors? Double-check the key and permissions.
• Use --headless=false to debug visually.

Resources to Explore

• Hugging Face Blog on Tiny Agents
• Playwright MCP GitHub
• TinyAgent GitHub
• Model Context Protocol Docs
• Chat Template Playground

Final Thoughts

You don’t need a massive framework to build a useful AI agent. Sometimes, all it takes is 50 lines of code and a well-defined task.

Tiny Agents give you everything you need to start building. Whether it’s scraping a website, reading a file, or automating a workflow—you can do it fast, and you can do it clean.

Ready to try it out? Fork the code, start your terminal, and give your AI a browser.

Written for The Data Career
Subscribe for more hands-on AI tutorials.