In this post, we will show you how to use Microsoft’s Open-Source Human-centric web agent, Magentic-UI. Magentic-UI has gained immense popularity among developers and researchers shortly after its release in May 2025. It’s being highlighted as a promising research prototype in agentic AI due to its “human-in-the-loop” transparency and safety features.
If you want to explore Magentic-UI, read this post. We’ll walk you through all the steps needed to install and use it on your Windows 11 or Windows 10 PC.
What is Magnetic UI?
Magentic-UI is a human-centric web agent developed by Microsoft that helps you perform complex, multi-step tasks through an interactive browser interface. You provide a task description, and it creates a plan; then, it sets up a virtual environment (like a mini-computer with a browser) to execute it, while keeping you in control.
How to use Microsoft’s Open-Source Human-centric web agent Magentic-UI?
In the following section, we will show you how to use Microsoft’s Open-Source Human-centric web agent Magentic-UI, 100% free, locally, using Ollama. But before that, let’s understand how it works.
How Magentic‑UI Works: A Team of Specialized AI Agents
Magentic-UI is powered by a team of intelligent agents adapted from AutoGen’s Magentic-One system (AutoGen is Microsoft’s open-source framework for building multi-agent AI systems). It consists of an Orchestrator agent, a WebSurfer agent, a Coder agent, and a FileSurfer agent.
- The Orchestrator is the lead agent. It builds a step-by-step plan, asks for user input when needed, and assigns subtasks to other agents.
- WebSurfer is a browser-controlling agent that can navigate websites, click, scroll, and upload files.
- Coder writes and executes Python and shell commands inside a secure Docker container to provide a response to the Orchestrator.
- FileSurfer reads and converts files using tools like MarkItDown and answers questions about them.
- A UserProxy agent also exists that represents the user interacting with Magentic-UI. It allows the system to assign specific steps back to the human user.
When a user gives a task, Magentic‑UI first generates a natural-language plan and presents it to the user. The user can add, delete, edit, regenerate steps, and write follow-up messages to iterate on the plan. The plan is stored inside the Orchestrator and is used to execute the task.
As each step is executed, the Orchestrator decides who should handle it: another agent or the user. The Orchestrator sends a request to one of the agents or the user and waits for a response. Once the response is received, the Orchestrator decides whether that step is complete. If the step is complete, the Orchestrator moves on to the next step.
Once all steps are completed, the Orchestrator compiles and presents the final result to the user.
Throughout the process, users can pause execution, adjust steps, or choose whether certain agent actions (like clicking buttons) require manual approval. This transparent and user-guided flow is what makes Magentic‑UI both powerful and safe to use.
Setting up Magentic‑UI with Ollama
To use Magentic-UI with local AI models (100% free and private), you do need to install and run Ollama on your Windows 11/10 PC.
Ollama is a free tool that lets you run large language models (LLMs) like llama3, mistral, or qwen2.5 directly on your computer, without using cloud services or API keys. Magentic-UI can connect to Ollama to use these models as the “brains” behind the agents.
Note: Before you begin, ensure you have admin rights (for initial installations) and your PC supports virtualization (required for Docker). Also, ensure you have at least 4–8 GB RAM and ~5–8 GB free disk space.
Step 1: Install Python
Go to www.python.org/downloads/ and download the latest Python 3.x version for Windows. Run the installer. During installation, check the box that says Add Python to PATH. Click Install Now.
Step 2: Install Docker Desktop
Go to www.docker.com/products/docker-desktop/ and download Docker Desktop for Windows. Install it and follow the setup wizard (use WSL 2 backend).
After installation, reboot your PC, open Docker Desktop, and ensure it’s running (you should see a little whale icon in the system tray).
Step 3: Install Ollama
Next, visit www.ollama.com and download Ollama for your system. Install it by running the downloaded installer.
Next, open Command Prompt and run the following command in it:
ollama run qwen2.5:3b
The above command will download the Qwen 2.5 3B model and start running it locally. You’ll see a waiting prompt like:
>>
Leave this window open. The model is now ready for Magentic‑UI to use.
Step 4: Install Magentic-UI
In a new Command Prompt window and run:
pip install "magentic-ui[ollama]"
The above command installs Magentic-UI and prepares your system to run it with local language models through Ollama. Next, run:
magentic-ui --port 8081
The above command will launch Magentic-UI locally on your system.
Note:
- The full setup process for Magentic‑UI may take 15–45 minutes, depending on your internet speed and system performance. Be patient during downloads and installations, especially when Docker or Ollama pulls large models. Once set up, everything runs locally and offline.
- It’s normal to see some red-colored lines or error-like messages in the terminal during the first-time startup of Magentic‑UI, especially when using Docker or Python environments. As long as the system continues running and doesn’t crash, you can safely ignore these warnings and temporary setup messages.
Next, open your browser and navigate to:
http://127.0.0.1:8081
You will see the local Magentic‑UI interface.
Step 5: Configure Magentic‑UI to use Ollama
In your browser, go to Settings > Agent Settings, and set the following:
- Provider: Ollama
- Model: qwen2.5:3b
Save the settings. You can now start using Magentic‑UI, powered by a local, free, private LLM running via Ollama.
Using Magentic-UI to perform web-based tasks
On top of the Magentic‑UI, you’ll see a text box where you can start putting your prompts. You can also attach files, get inspired by the example prompts, and run multiple sessions in parallel.
Once you submit a task, Magentic‑UI shows a step-by-step draft plan in the center and a Live View panel on the right. This panel displays a simulated browser view showing the agent’s real-time activity (clicks, typing, scrolling, tab switching, etc.), along with screenshots (of pages visited) and brief logs. There’s also a Take Control button that allows you to pause the agent and step in to directly interact with the virtual browser.
In the plan, Magentic-UI shows the steps to execute the task. Each step shows a small delete button beside it when you hover over it. This helps remove the step from the sequence. If you want to edit a step, click on it to make it editable. Besides this, you can reorder steps using drag-and-drop or insert a new step manually by clicking the Add Step button at the bottom.
After reviewing the plan, you can either click Accept Plan to proceed with execution or Generate New Plan if you want the agent to create a new version of the steps automatically.
Once you click Accept Plan, Magentic‑UI begins executing each step using its agents. You’ll see live progress in the right panel, including browser actions, screenshots, and logs. You can pause execution, take control, or provide feedback at any point during the process.
You can read more about this on Microsoft and GitHub.
Read: How to choose the best LLM for your Task?
What is the use of magnetic one?
The main purpose of Magentic-One is to coordinate multiple specialized agents (like a browser agent, coding agent, or file-handling agent) so they can work together on complex, multi-step tasks, with or without human guidance. It is designed to make AI agents smarter and more collaborative by organizing them into a structured, task-sharing system.
Read Next: How to build AI Agent: Beginners Guide from Microsoft.
