mcp-data-summary

A learning project for building an MCP server with the official Python SDK.

The server lets an LLM produce a visual summary report from CSV datasets without ever loading raw data into the model's context. It simulates a REST API (fake GET/POST), exposes data schemas as MCP Resources, provides chart and document generation as MCP Tools, and guides the whole workflow with an MCP Prompt.

---

What it demonstrates

MCP concept	File	What it does
Resources	resources/datasets.py	Expose dataset schemas via fake GET endpoints
Tools	tools/chart_tools.py	Generate bar, line, histogram, pie charts, dataset subset
Tools	tools/stats_tools.py	Descriptive stats + correlation matrix + scatter plot
Tools	tools/document_tools.py	Build HTML report + export PDF via fake POST
Prompts	prompts/summary_prompt.py	7-step guided workflow for the LLM
Fake API	api/fake_client.py	Simulated REST client backed by local CSV and JSON files
Schemas	api/schemas.py	Pydantic models for all API responses

---

How it works

The LLM never sees raw tabular data. Instead it follows this pipeline:

GET /datasets                 → discover available datasets
GET /datasets/{name}          → inspect a plot-friendly schema (columns, dtypes, sample values)
tool: get_summary_statistics  → understand distributions and correlations
tool: generate_*_chart        → produce PNG charts saved to output/ and feedbacks to understand the charts
tool: build_html_report       → render a Jinja2 HTML report from the chart paths
tool: export_pdf_report       → convert HTML to PDF and POST it back to the fake API

---

Project structure

mcp-data-summary/
├── data/                              # Sample CSV files and JSON dataset descriptions (auto-discovered)
│   ├── descriptions.json
│   ├── sales.csv
│   └── users.csv
├── output/                            # Generated charts, HTML and PDF reports
├── scripts/
│   └── run_pipeline.py                # End-to-end pipeline runner (no LLM needed)
├── tests/
│   ├── conftest.py                    # Shared pytest fixtures
│   ├── test_fake_client.py            # Unit tests for the API layer
│   ├── test_chart_tools.py            # Tests for all chart tools
│   └── test_stats_and_docs.py         # Tests for stats tool + HTML builder
├── src/
│   └── mcp_data_summary/
│       ├── server.py                  # Entry point — wires everything together
│       ├── api/
│       │   ├── fake_client.py         # Simulates GET /datasets and POST /reports
|       |   ├── json_encoder.py        # Custom JSON encoder
│       │   └── schemas.py             # Pydantic models for API responses
│       ├── resources/
│       │   └── datasets.py            # MCP Resources: dataset list + schema
│       ├── tools/
│       │   ├── chart_tools.py         # Bar, line, histogram, pie chart tools
│       │   ├── stats_tools.py         # Summary statistics + scatter plot
│       │   └── document_tools.py      # HTML report builder + PDF exporter
│       └── prompts/
│           └── summary_prompt.py      # Guided 7-step workflow prompt
└── pyproject.toml

---

Quick start

1. Install uv

curl -LsSf https://astral.sh/uv/install.sh | sh

2. Install dependencies

uv sync

3. Verify everything works (no LLM required)

Run the pipeline script — it simulates the full LLM workflow end-to-end:

uv run python scripts/run_pipeline.py

# Restrict to a single dataset
uv run python scripts/run_pipeline.py --datasets sales

Charts and a report will appear in output/.

4. Run the test suite

uv run pytest

5. Open the MCP Inspector (browser-based debug UI)

uv run mcp dev src/mcp_data_summary/server.py

This lets you browse Resources, call Tools manually, and inspect request/response payloads

6. Example - Connect to Claude Desktop

Add this block to your claude_desktop_config.json (~/Library/Application Support/Claude/ on macOS, %APPDATA%\Claude\ on Windows):

{
  "mcpServers": {
    "data-summary": {
      "command": "uv",
      "args": [
        "--directory", "/absolute/path/to/mcp-data-summary",
        "run", "mcp-data-summary"
      ]
    }
  }
}

Restart Claude Desktop, then load the data_summary_workflow prompt. The LLM will discover datasets, inspect schemas, generate charts, and produce a PDF report autonomously.

---

Available tools

Tool	Input	Output
get_summary_statistics	dataset name	JSON with describe + correlation
generate_bar_chart	dataset, x/y columns, optional group_by	PNG path and chart data
generate_line_chart	dataset, x/y columns, optional group_by	PNG path and chart data
generate_histogram	dataset, column, bins	PNG path and chart data
generate_pie_chart	dataset, category + value columns	PNG path and chart data
generate_scatter_plot	dataset, x/y columns, optional color_by	PNG path and chart data
build_subset_dataset	dataset, filters	Subset name, path and schema
build_html_report	title, chart paths, captions, summary	HTML path
export_pdf_report	HTML path, report name	JSON with pdf path + report_id

---

Available resources

URI	Returns
datasets://list	JSON array of available dataset names
datasets://{name}/schema	Plot-friendly schema: columns, dtypes, n_unique, sample values

---

Adding your own CSV

Drop any .csv file into data/. The server auto-discovers it on startup. Add a description in data/descriptions.json:

{
    "your_file": "Description the LLM will see when reading the schema.",
}

Date columns are auto-detected if their name contains "date".

---

System dependencies for PDF export

WeasyPrint (used by export_pdf_report) requires Cairo and Pango to be installed at the OS level. The HTML report and all charts work without them.

# Ubuntu / Debian
sudo apt install libpango-1.0-0 libcairo2 libpangocairo-1.0-0

# macOS
brew install pango cairo

# Windows — follow the WeasyPrint install guide:
# https://doc.courtbouillon.org/weasyprint/stable/first_steps.html

---

Next steps

• Support multi-dataset joins before charting
• Introduce a simple auth token to the fake API layer
• Stream chart generation progress back to the client using MCP notifications