panel-live-server¶

Building interactive data apps with Panel, hvPlot, or HoloViews usually means: write code, start a server, open a browser, refresh, repeat. That loop adds up for quick experiments. And if you are working inside an AI assistant like Claude or GitHub Copilot, asking it to help you build a chart, there has been no clean way to actually see the result without leaving your tool.

panel-live-server removes that friction. Point an MCP-compatible AI assistant at a dataset or describe the chart you want, and the result comes back rendered, live, and interactive, right inside the chat. Or run it standalone from the terminal: submit code through a browser UI and get a permanent URL back instantly.

It ships two interfaces, built on the same underlying Panel server:

• MCP server (pls mcp): connects to AI assistants over the Model Context Protocol, so they can render and inspect visualizations directly inside the chat
• Standalone server (pls serve): a web server you drive yourself, through a browser UI or REST API

Use whichever fits how you work, or run both.

MCP Server: AI assistant integration¶

Give Claude, GitHub Copilot, Cursor, or any MCP-compatible AI assistant the ability to render visualizations directly in your IDE, and to actually see what it just rendered. Two tools are exposed:

• show: validates the code (syntax, security, package availability, Panel extensions) and then executes it, returning a live, interactive visualization — no manual setup and no separate validation step required. The AI is instructed to reach for HoloViz packages (hvPlot, HoloViews, Panel) first, falling back to other well-known libraries only when needed
• screenshot: captures a picture of an already-rendered visualization and hands it back to the AI, so it can answer follow-up questions about how the chart looks by inspecting the actual image instead of guessing from raw data

Install the package, then start the MCP server:

pixi add --pypi "panel-live-server[pydata]"

uv tool install "panel-live-server[pydata]"

pip install "panel-live-server[pydata]"

pls mcp  # configure this command in Claude, Copilot, etc.

See the Installation tutorial for per-package-manager setup details and connecting to your MCP client.

Ask your AI assistant:

Please show a quick and beautiful Matplotlib trading dashboard

Please show a basic, interactive Panel app with a slider.

Now replace the text with a hvplot and show it.

Please show the most beautiful matplotlib plot

The AI calls show to render it, and the visualization appears immediately in your chat interface. If you then ask a follow-up question about how it looks, the AI can call screenshot to look at the rendered image before answering.

Standalone Server: browser UI and REST API¶

Start a local web server and create interactive visualizations through a browser UI or REST API. Every snippet gets its own permanent URL.

Install the package, then start the server:

pixi add --pypi "panel-live-server[pydata]"

uv tool install "panel-live-server[pydata]"

pip install "panel-live-server[pydata]"

pls serve  # run this command in the terminal

Open http://localhost:5077/add and submit any Python visualization:

import pandas as pd
import hvplot.pandas

df = pd.DataFrame({'Product': ['A', 'B', 'C', 'D'], 'Sales': [120, 95, 180, 150]})
df.hvplot.bar(x='Product', y='Sales', title='Sales by Product')

Browse your visualizations at /feed, manage them at /admin, and link directly to any individual chart at /view?id=....

Features¶

Two execution methods¶

• Inline (default): the last expression is automatically displayed, just like a notebook cell
• Server: explicit .servable() calls for multi-component dashboards with reactive widgets

Works with any Python visualization library¶

hvplot · plotly · altair · matplotlib · seaborn · holoviews · bokeh · vega · deckgl · and more

Persistent storage¶

Every snippet is saved to a local SQLite database with full-text search. Visualizations survive server restarts and are accessible by URL at any time.

Robust subprocess management¶

The Panel server runs as a managed subprocess with health monitoring and automatic restart (up to a configurable limit). Port conflicts and stale processes are handled automatically.

Validate before you render¶

show runs four static checks (syntax, security, package availability, and Panel extension declarations) automatically before it executes anything. Validation is built into the render path, so there is no separate step to call and no double-validation overhead.

See it, don't guess¶

A screenshot tool captures a PNG of an already-rendered visualization in a headless browser and hands it to the AI. This lets the assistant correctly answer questions like "which bar is tallest?" or "where does it peak?" by looking at the actual rendered output, since plots routinely flip axes, reorder rows, or bin values differently than the raw data suggests.

MCP App UI¶

When used with a compatible AI client, visualizations render inline with zoom controls (25 / 50 / 75 / 100 %), one-click URL and code copying, and a loading indicator.

REST API¶

import requests

response = requests.post(
    "http://localhost:5077/api/snippet",
    json={"code": "1 + 1", "name": "Addition", "method": "inline"}
)
print(response.json()["url"])

Works everywhere¶

Local, Jupyter, JupyterHub, VS Code Dev Containers, GitHub Codespaces: URLs are automatically externalized via Jupyter Server Proxy when needed.

Learn more¶