panel-viz-mcp

Interactive Panel / HoloViews visualizations rendered directly inside AI chat UIs via the MCP Apps Standard.

Built with FastMCP · hvPlot · Bokeh · Panel

14 chart types · 15 MCP tools · 4 interactive UI resources · bidirectional communication

Showcase

Inline Charts in AI Chat

Type a prompt, get an interactive chart rendered directly in the conversation.

Live Streaming

Real-time data simulation with play/pause controls - all rendered inline.

Multi-Chart Grid

Side-by-side chart comparison from the same dataset.

Geographic Maps

Interactive maps with tile basemaps, launched via "Open in Panel".

Candlestick Portfolio Dashboard

Full OHLC candlestick chart with donut pie breakdown, styled Tabulator table with colored buy/sell/hold signals, Number indicators, and crossfiltering sidebar - all generated from a single prompt.

Medical Image Viewer

Real brain MRI data (scikit-image), 2x2 grid with multiple colormaps, slice navigation, brightness/contrast controls - built via create_panel_app.

ML Model Evaluator

Interactive ML dashboard with confusion matrix, ROC curve, feature importances, classification report, and a "Retrain Model" button that retrains a RandomForest with new hyperparameters and updates all 4 panels live.

9 Demo Prompts - Ready-to-use prompts with full data included. From simple bar charts to candlestick portfolio dashboards, medical image viewers, and ML model evaluators. Just copy, paste, and run.

What is this?

An MCP server that lets AI assistants create, modify, and render interactive visualizations inline in the chat conversation. Charts render as live BokehJS figures inside sandboxed iframes - not static images, not external links.

Works with any MCP Apps-compatible client:

VS Code Copilot Chat (recommended - best inline iframe rendering)
Claude Desktop / Claude Code
Cursor
ChatGPT (Business/Enterprise/Edu)
Goose

Quick Start

Install

git clone https://github.com/AtharvaJaiswal005/panel-viz-mcp.git
cd panel-viz-mcp
pip install -e .

For geographic maps: pip install -e ".[geo]"

Connect to your client

VS Code / Copilot Chat (Recommended)

Create .vscode/mcp.json in your project folder:

{
  "servers": {
    "panel-viz-mcp": {
      "command": "panel-viz-mcp"
    }
  }
}

Open Copilot Chat (Ctrl+Shift+I), switch to Agent mode, and you're ready.

Claude Code

claude mcp add panel-viz-mcp -- panel-viz-mcp

Claude Desktop

Open Settings > Developer > Edit Config and add:

{
  "mcpServers": {
    "panel-viz-mcp": {
      "command": "panel-viz-mcp"
    }
  }
}

Save and restart Claude Desktop.

Cursor / Goose / Other Clients

Use stdio transport with command panel-viz-mcp.

Or let the agent install it for you

Paste one prompt and the AI agent handles everything - clone, install, configure, verify.

VS Code Copilot Chat prompt

Install the panel-viz-mcp MCP server for this workspace.
1. Clone: git clone https://github.com/AtharvaJaiswal005/panel-viz-mcp.git
2. Install: cd panel-viz-mcp && pip install -e .
3. Create .vscode/mcp.json with the server command set to "panel-viz-mcp"
4. Confirm the server is connected.

Claude Code prompt

Install the panel-viz-mcp MCP server.
1. Clone: git clone https://github.com/AtharvaJaiswal005/panel-viz-mcp.git
2. Find my Python path (run "where python" on Windows or "which python3" on Mac/Linux)
3. Install: cd panel-viz-mcp && pip install -e .
4. Register: claude mcp add panel-viz-mcp -- panel-viz-mcp
5. Confirm the server is registered.

Claude Desktop prompt

Install the panel-viz-mcp MCP server.
1. Clone: git clone https://github.com/AtharvaJaiswal005/panel-viz-mcp.git
2. Find my Python path (run "where python" on Windows or "which python3" on Mac/Linux)
3. Install: cd panel-viz-mcp && pip install -e .
4. Add to claude_desktop_config.json with the Python path as command and ["-m", "panel_viz_mcp.server"] as args
5. Tell me to restart Claude Desktop when done.

MRVE - Minimum Reproducible Viable Example

Want to understand the core concept without the full 15-tool server? mrve.py is a single-file (~200 lines) demo showing:

Inline rendering: LLM calls create_viz -> hvPlot builds chart -> Bokeh json_item() -> BokehJS renders inside MCP Apps iframe
Bidirectional communication: User clicks chart -> iframe calls handle_click via callServerTool() -> server computes insight -> UI displays it

fastmcp run mrve.py

2 tools (create_viz, handle_click), 1 HTML resource, 4 chart types (bar, line, scatter, area). That's it.

Try These Prompts

Just paste any of these into your AI chat.

Prompt	What you get
`Create a bar chart showing quarterly revenue: Q1: 42000, Q2: 58000, Q3: 71000, Q4: 89000`	Interactive bar chart inline with hover tooltips and click insights
`Create a dashboard of this sales data with filters: Region, Product, Sales...`	Full dashboard with chart + stats + table + working filter sidebar
`Create a live streaming stock price chart starting at $150`	Real-time updating chart with play/pause/reset controls
`Show me a bar chart and scatter plot of this data side by side`	Multi-chart grid layout from the same dataset
`Create a geographic points map of India's major cities...`	Geographic map with tile basemaps (opens in Panel)
`Create a candlestick chart for RELIANCE stock with this OHLC data...`	Candlestick chart with green/red candles (opens in Panel)

See all 9 demo prompts with full data - includes portfolio dashboards, medical image viewers, and ML model evaluators.

Features

14 Chart Types

Type	Description	Special Parameters
`bar`	Vertical bar chart	-
`line`	Line chart	-
`scatter`	Scatter plot	-
`area`	Filled area chart	-
`pie`	Pie/donut chart	-
`histogram`	Frequency distribution	-
`box`	Box plot	`x` = grouping column (categorical), `y` = numeric
`violin`	Violin plot	`x` = grouping column (categorical), `y` = numeric
`kde`	Kernel density estimate	Only `y` needed (density of that column)
`step`	Step chart	-
`heatmap`	Categorical heatmap	`color` = value column (C)
`hexbin`	Hexagonal binning	Both `x` and `y` must be numeric
`points`	Geographic scatter map	`x` = longitude, `y` = latitude, requires `geoviews`
`candlestick`	OHLC candlestick chart	Data must have Open, High, Low, Close columns

15 MCP Tools

#	Tool	Description
1	`create_viz`	Create a chart (any of 14 types)
2	`update_viz`	Modify an existing chart (change type, data, axes, title)
3	`load_data`	Load CSV/Parquet/JSON/Excel/Feather/Zarr and visualize
4	`handle_click`	Process click events from charts (bidirectional)
5	`list_vizs`	List all active visualizations
6	`create_dashboard`	Dashboard with chart + stats + table + filter widgets
7	`stream_data`	Live-updating streaming chart
8	`apply_filter`	Apply widget filters to a dashboard (crossfiltering)
9	`set_theme`	Switch between dark and light theme
10	`create_multi_chart`	2-4 chart grid layout from the same dataset
11	`annotate_viz`	Add annotations (hline, vline, text, band, arrow)
12	`export_data`	Export data as CSV or JSON
13	`launch_panel`	Open a full interactive Panel app in the browser
14	`stop_panel`	Stop a running Panel server
15	`create_panel_app`	Launch a custom Panel app from LLM-written Python code

4 Interactive UI Resources (MCP Apps)

Resource	Description
`ui://panel-viz-mcp/viz.html`	Single chart viewer with toolbar (theme toggle, Save PNG, Export CSV, Open in Panel)
`ui://panel-viz-mcp/dashboard.html`	Dashboard with chart, summary stats, data table, and filter sidebar
`ui://panel-viz-mcp/stream.html`	Live streaming chart with play/pause/reset controls
`ui://panel-viz-mcp/multi.html`	Multi-chart grid (2-4 charts from the same data)

Key Capabilities

Bidirectional communication - click a data point in the chart, get AI-generated insights back
Server-side crossfiltering - filter sidebar with Select dropdowns and Range sliders, filters update chart + stats + table
Live streaming - real-time data simulation with client-side BokehJS updates
Annotations - horizontal/vertical lines, text labels, bands, arrows
Theme switching - dark/light mode toggle with CSS variables + Bokeh re-rendering
"Open in Panel" button - launch a full interactive Panel app in the browser from any inline chart
Rich Panel apps - generated apps include FloatPanel chart inspector, Tabulator data table, Number indicators, crossfiltering sidebar, FastListTemplate dark theme
Save PNG / Export CSV - toolbar buttons on every chart

Architecture

User prompt
  -> AI calls MCP tool (e.g. create_viz)
    -> Python builds chart via hvPlot / HoloViews
      -> hv.render(backend="bokeh") converts to Bokeh figure
        -> bokeh.embed.json_item(fig) serializes to JSON
          -> JSON sent to MCP Apps iframe
            -> BokehJS renders interactive chart inline in chat UI

Bidirectional flow:

User clicks data point in chart
  -> BokehJS TapTool fires callback
    -> app.callServerTool("handle_click", {viz_id, point_index, x_value, y_value})
      -> Server generates insight text
        -> Response displayed in chart's status bar

Filter flow:

User changes filter widget in dashboard
  -> app.callServerTool("apply_filter", {viz_id, filters})
    -> Server filters DataFrame, rebuilds chart + stats + table
      -> Updated figure JSON + stats + table HTML returned
        -> Dashboard re-renders chart, stats panel, and data table

Development

git clone https://github.com/AtharvaJaiswal005/panel-viz-mcp.git
cd panel-viz-mcp
pip install -e ".[dev]"

Run tests

pytest tests/ -v

51 tests cover all 15 tools, 14 chart types, edge cases, code generation, and security sandboxing.

Project structure

panel-viz-mcp/
  src/panel_viz_mcp/
    __init__.py          # Package re-export
    app.py               # FastMCP instance + in-memory stores
    server.py            # Thin entry point with re-exports
    constants.py         # Chart types, URIs, limits, SDK version
    cdn.py               # BokehJS CDN script tags
    themes.py            # Theme colors + CSS variables
    chart_builders.py    # Bokeh figure construction + annotations
    code_generators/     # Panel app code generators
      standard.py        # Single-chart Panel apps
      geo.py             # DeckGL geographic apps
      multi.py           # Multi-chart grid apps
    tools/               # 15 MCP tools
      viz.py             # create, update, load, click, list
      dashboard.py       # create_dashboard, apply_filter, set_theme
      stream.py          # stream_data
      multi.py           # create_multi_chart
      annotation.py      # annotate_viz
      export.py          # export_data
      panel_launch.py    # launch_panel, stop_panel
      custom_app.py      # create_panel_app
    resources/           # 4 HTML resources (MCP Apps)
      viz_html.py        # Single chart viewer
      dashboard_html.py  # Dashboard with filters
      stream_html.py     # Live streaming chart
      multi_html.py      # Multi-chart grid
  tests/
    test_tools.py        # 51 pytest tests
  demos/
    DEMO_SCRIPT.md       # 9 copy-paste demo prompts
    ml_dashboard.py      # Standalone ML dashboard example
  examples/
    sample_data.csv      # Sample CSV for load_data tool
  docs/
    screenshots/         # GIFs and screenshots for README
  pyproject.toml
  LICENSE                # MIT
  README.md

Tech Stack

FastMCP 3.x - MCP server framework with Apps support
Panel - Interactive web apps and dashboards
HoloViews - Declarative data visualization
hvPlot - High-level plotting API for pandas
Bokeh - Interactive visualization library (rendering backend)
GeoViews - Geographic visualizations (optional, for points type)

Related Projects

holoviz-mcp by Marc Skov Madsen - MCP server for HoloViz documentation and code execution
panel-live - Run Panel code in the browser with MCP support
Panel #8396 - Tracking issue for MCP Apps Standard support in Panel core

License

MIT

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

panel-viz-mcp

Showcase

Inline Charts in AI Chat

Live Streaming

Multi-Chart Grid

Geographic Maps

Candlestick Portfolio Dashboard

Medical Image Viewer

ML Model Evaluator

What is this?

Quick Start

Install

Connect to your client

Or let the agent install it for you

MRVE - Minimum Reproducible Viable Example

Try These Prompts

Features

14 Chart Types

15 MCP Tools

4 Interactive UI Resources (MCP Apps)

Key Capabilities

Architecture

Development

Run tests

Project structure

Tech Stack

Related Projects

License

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 36 Commits
demos		demos
docs/screenshots		docs/screenshots
examples		examples
src/panel_viz_mcp		src/panel_viz_mcp
tests		tests
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
mrve.py		mrve.py
pyproject.toml		pyproject.toml

Folders and files

Latest commit

History

Repository files navigation

panel-viz-mcp

Showcase

Inline Charts in AI Chat

Live Streaming

Multi-Chart Grid

Geographic Maps

Candlestick Portfolio Dashboard

Medical Image Viewer

ML Model Evaluator

What is this?

Quick Start

Install

Connect to your client

Or let the agent install it for you

MRVE - Minimum Reproducible Viable Example

Try These Prompts

Features

14 Chart Types

15 MCP Tools

4 Interactive UI Resources (MCP Apps)

Key Capabilities

Architecture

Development

Run tests

Project structure

Tech Stack

Related Projects

License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages