New: Dash MCP Server

Dash Python
1

:rocket: Try the new Dash MCP server with pip install dash==4.3.0rc0

I’ve recorded one video showing the usefulness of Dash MCP and another video showing how to incorporate Dash MCP as a connector in Claude.

The Dash MCP server exposes your app’s callbacks to the LLMs as tools, allowing your AI agent to inspect layouts, read component states, and iterate, all through natural language.

On June 23 we will host a webinar that walks you through a live demo of Dash MCP, build a complete Dash app from the ground up with Claude, and discuss additional use cases.

:backhand_index_pointing_right: Register for the Webinar!

Quick Start with the Dash 4.3.0 release candidate and Dash MCP

app = Dash(__name__, enable_mcp=True)

Then add your app’s URL as an MCP server in your AI agent.

image
image1024×1024 170 KB

Dash() constructor parameters

Parameter Type Default Description
enable_mcp bool False Turns the whole MCP system on/off
mcp_path str /_mcp Route the MCP server is served from
mcp_expose_docstrings bool False Expose all callback docstrings to agents

Also settable via env vars: DASH_MCP_ENABLED, DASH_MCP_PATH, DASH_MCP_EXPOSE_DOCSTRINGS (constructor wins over env).

Callbacks as tools (default behavior)

All callbacks are exposed as MCP tools automatically. The server describes each callback’s Inputs/Outputs/State and how they tie to layout (component types, allowed values, defaults, labels). Per-callback controls are also an option, instead of enabling globally:

@callback parameter Type Default Description
mcp_enabled bool True Set False to hide a callback from agents
mcp_expose_docstring bool False Include this callback’s docstring as agent context
  • Type annotations (e.g. value: List[int]) narrow the type agents send, avoiding trial-and-error.
  • Special return values: a Plotly figure is rendered to PNG via Kaleido (if installed); a DataFrame is rendered as a Markdown table — for agents that support images/tables.

Custom tools — @mcp_enabled decorator

Expose any plain Python function (not just callbacks) as a tool:

from dash.mcp import mcp_enabled

@mcp_enabled
def get_sales(region: str) -> dict: ...

@mcp_enabled(name="get_sales", expose_docstring=True)
def get_sales_by_region(region: str) -> dict:
    """Return sales for a region (north, south, east, west)."""
    ...
Argument Type Default Description
name str function name Tool name shown to agents/users
expose_docstring bool False Include the docstring in the tool description

Authentication

Dash implements the MCP server but not auth. When you deploy your app to Plotly Cloud, auth is configured automatically, and only users with access to your app can access the MCP server. Get started with Plotly Cloud for free.

Security notes

  • Source code and private variables are not exposed — only an AI-friendly view of layout and callbacks.
  • Callback function names are exposed (agents show them when requesting permission), so avoid leaking internals in names.
  • MCP makes “security by obscurity” inputs trivially discoverable (though this risk already exists via browser/curl).
2

Hey @adamschroeder

That’s interesting. So the MCP server is for the specific app only? What are you referring to if you mention “claude has access to the code” in one of your videos?

I would assume this is for development only or would you activate the MCP for apps in production too?

3

Hi @AIMPED ,
so technically by adding the app-url.com/_mcp as a connector to Claude, it gives Claude access to the app that it can interact with in order to get more information from it. For example in my app, if you ask it about the number of Montreal Metro incidents in 2020, it would be able to use the app controls/callbacks to filter the dropdown to 2020 and see the data.

This should work for any Dash app on Plotly Cloud that is shared publicly, and which has: app=Dash(enable_mcp=True). By next week, the MCP server will work for private apps on Cloud as well.
In fact, the MCP server is meant to be used in production rather than dev, since it’s a feature that allows you to make your data app accessible to AI.

4

Great questions @AIMPED ! You can see in the updated description an explainer of the new API for MCP. You’re correct that it’s for the specific app only.

Claude won’t have access to the code directly, just the callbacks available as MCP tools.

For production deployments, we recommend Plotly Cloud, where authentication is handled automatically (only users with access to your app can use it as an MCP server).