anywidget-mcp 0.0.2

Expose AnyWidget interfaces as interactive MCP Apps

anywidget-mcp runs AnyWidgets inside AI conversations. You can interact with a widget directly, and the model can respond to your input.

Start with a widget from an existing package, or bring the same widget you use in Jupyter or marimo.

Read the documentation.

Try an AnyWidget

Requires Python 3.11 or newer. Start Wigglystuff's ColorPicker with one command:

uvx --with wigglystuff anywidget-mcp serve wigglystuff:ColorPicker --port 8010

Connect a host that supports MCP Apps to http://127.0.0.1:8010/mcp, then ask:

Use color_picker so I can choose a color.

The picker opens in the conversation. Choose a color, then ask the model which color you chose. The model reads the picker's current color when it answers.

Getting started walks through this flow in Inspector Chat.

But why?

💬 Stay in context: A standalone web app sends you to a separate page. An MCP App stays in the conversation, exchanges data through MCP, and runs in a host-controlled sandbox. With your consent, it can ask the host to use tools you already connected.

🧩 Skip the scaffolding: Building an MCP App directly means wiring together tools, UI resources, browser code, the host connection, and shared state. anywidget-mcp handles that integration. You define an AnyWidget that keeps browser behavior and Python state in one component.

🧪 Develop in notebooks: The same widget runs in Jupyter and marimo. Create it from Python, inspect its state, and use marimo's reactive execution to test scenarios before serving it through MCP.

🧰 Combine widgets: One server can expose multiple widgets as separate tools, or one call can open several together. Combine widgets from the AnyWidget gallery into an interactive environment for the task, then add your own.

✨ Invent at runtime: When you cannot enumerate every useful interface in advance, serve create_anywidget. Supply Python source directly or ask an agent to invent a fresh AnyWidget during the conversation. Use that natural-language loop to develop widgets and test concepts. Run the factory in a sandbox because supplied source executes with the MCP server's permissions.

Install in a project

Install anywidget-mcp in the Python environment that owns your widget code:

uv pip install anywidget-mcp

Bring your own AnyWidget

Point the CLI at your AnyWidget's import path:

anywidget-mcp serve my_widgets:MyWidget

Or start the server from Python:

from anywidget_mcp import serve
from my_widgets import MyWidget

serve(MyWidget)

MyWidget remains the same anywidget.AnyWidget subclass you use in notebooks. Browser changes stay synchronized with its Python traits.

When the widget needs data from the conversation, use a Python function that accepts the input and returns an AnyWidget. This function is a widget factory. Pass input to widgets covers that workflow, multiple widgets, and existing MCP servers.

Documentation

• Getting started
• Model-visible state
• Pass input to widgets
• How it works
• API reference
• Deployment