ARTPARK Public Data MCP Server

Devin powered Deepwiki Documentation

MCP (Model Context Protocol) server for accessing ARTPARK's public health datasets through AI assistants. Built with FastMCP 3.0, following the architecture of esankhyiki-mcp.

8 datasets covering Dengue epidemiology, Livestock census, FMD vaccination & surveillance, and administrative geospatial data from ARTPARK DSIH.

Quick Start

# Clone with submodules
git clone --recurse-submodules <repo-url>
cd onehealth-artpark-mcp

# Create environment and install
uv venv .venv && source .venv/bin/activate
uv pip install -r requirements.txt

# Run
python artpark_server.py
# → MCP:    http://localhost:8000/mcp
# → Health: http://localhost:8000/health

Or with Make:

make run          # Start MCP server
make test         # Run 40 tests (~2s)
make check        # Verify datasets + tools load
make docker-up    # Start server + Jaeger tracing
make help         # Show all commands

Related MCP server: CKAN MCP Server

Datasets

Data source: dsih-artpark/publicdata | Docs: publicdata.readthedocs.io

MCP Tools (4-Step Workflow)

1_know_about_artpark_data  →  2_get_tables  →  3_get_metadata  →  4_get_data

Important: Tools must be called in order. Skipping 3_get_metadata means you won't know valid column names or filter values.

What Can an AI Do With This?

The examples/Complex/ folder contains five investigations that were executed entirely by an AI using this MCP server -- no human analyst, no Jupyter notebooks, no code written. Each one demonstrates a capability that goes beyond data retrieval:

Read the full series README for the narrative arc connecting all five.

How We Generated These Reports

Every example in this repo was produced by giving Cursor's AI agent access to this MCP server and asking interesting questions. No Jupyter notebooks. No Python scripts. No human data wrangling. The AI used the 4-step MCP workflow autonomously, pulled real data, did the analysis in-context, and wrote the markdown reports.

Here's how to reproduce this -- or create your own investigations:

Step 1: Start the MCP server

make run
# or: python artpark_server.py

Step 2: Connect Cursor to the MCP server

Create (or verify) .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "artpark": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Restart Cursor so it picks up the MCP config. You should see artpark listed under the MCP tools in the agent chat.

Step 3: Ask the AI a question

Open Cursor's agent chat and ask something interesting. The simple examples (01-04) came from single questions like:

• "Is India's FMD vaccination program working?" → Cross-Dataset Reasoning

• "Tell me the story of Karnataka's FMD vaccination journey." → Temporal Narrative

• "Any surprising differences in FMD across livestock types?" → Hidden Patterns

• "Where should Karnataka prioritize its next FMD vaccination?" → Policy Intelligence

Step 4: Go bigger -- ask for ambitious investigations

The Complex investigations were generated in a multi-step conversation:

1. Brainstorm: We asked: "Think of some really ambitious ideas -- can we build a model? Find a needle in a haystack? Debunk a common myth?" The AI generated 5 investigation ideas with detailed instructions.

2. Execute: We said: "Build each one of them. You'll interact with the MCP server and report back what you find." The AI then autonomously:

   • Called the MCP tools to pull data from multiple datasets

   • Cross-referenced datasets (e.g., joining dengue + livestock data across 17 districts)

   • Computed statistics (Spearman correlations, immunity crash deltas, risk scores)

   • Wrote full markdown reports with data tables, methodology, and conclusions

3. Refine: We prompted "go on" and "refine it" several times -- each investigation required multiple agent turns to complete the analysis and polish the report.

4. Verify: We asked the AI to spot-check its own numbers against the live MCP data to catch any hallucinated figures.

Step 5: Try your own questions

Some ideas to get you started:

• "Which Karnataka district has the worst dengue-to-healthcare ratio?"

• "Compare FMD vaccination schedules vs actual progress -- which rounds fell behind?"

• "Find states where natural FMD infection is rising despite vaccination."

• "Build a composite One Health risk index for Karnataka districts."

The AI will follow the 1_know → 2_get_tables → 3_get_metadata → 4_get_data workflow automatically. Responses will vary (LLMs are non-deterministic), but the cross-dataset reasoning and findings are reproducible.

Tip: For long investigations, tell the AI "You may require long runs -- feel free to not finish in one call and ask for more time." This lets it pace itself instead of rushing through a shallow analysis.

Quick Examples

1. Dengue Outbreak Analysis in Karnataka

"How many dengue cases were reported in Bengaluru Urban?"

# Step 1: Find the right dataset
overview = await client.call_tool("1_know_about_artpark_data", {})
# → Dataset 0015: "Karnataka Dengue Daily Summary (2017-2024)"

# Step 2: See what tables are available
tables = await client.call_tool("2_get_tables", {
    "dataset_id": "0015",
    "user_query": "dengue cases by district"
})
# → Table: ka-dengue-daily-summary (28 columns)

# Step 3: Get column schema and filter values
meta = await client.call_tool("3_get_metadata", {
    "dataset_id": "0015",
    "table_name": "ka-dengue-daily-summary"
})
# → 76,114 rows, columns include daily.suspected, daily.positive.total, daily.deaths

# Step 4: Fetch Bengaluru data
data = await client.call_tool("4_get_data", {
    "dataset_id": "0015",
    "table_name": "ka-dengue-daily-summary",
    "filters": {"location.admin2.name": "Bengaluru Urban"},
    "limit": 50
})
# → 4,942 records, positive cases: min=0, max=1070, mean=7.49

2. Livestock Population Comparison Across Districts

"Which Karnataka districts have the most cattle?"

data = await client.call_tool("4_get_data", {
    "dataset_id": "0041",
    "table_name": "ka-district-livestock-pop-2019",
    "limit": 30
})
# Results (top 5 by cattle population):
#   BELAGAVI:   cattle=549,540  buffalo=844,171  sheep=757,679
#   HASSAN:     cattle=548,185  buffalo=107,971  sheep=199,387
#   SHIVAMOGGA: cattle=518,653  buffalo=120,563  sheep=42,526
#   MYSURU:     cattle=492,598  buffalo=21,682   sheep=203,463
#   TUMAKURU:   cattle=431,251  buffalo=142,047  sheep=1,290,008

3. FMD Vaccine Efficacy Tracking

"How has FMD seromonitoring changed in Karnataka over the years?"

data = await client.call_tool("4_get_data", {
    "dataset_id": "0087",
    "table_name": "seromonitoring",
    "filters": {"state.name": "KARNATAKA"},
    "limit": 20
})
# Results (antibody levels over time):
#   2011: PostVac A%=15.0, PostVac O%=56.0  (FMDCP)
#   2012: PostVac A%=27.0, PostVac O%=67.0  (FMDCP)
#   2013: PostVac A%=78.7, PostVac O%=62.1  (FMDCP)
#   → Shows improving seroconversion rates for type A

4. Administrative Region Lookup

"Look up LGD codes for Indian administrative units"

data = await client.call_tool("4_get_data", {
    "dataset_id": "0034",
    "table_name": "regionids",
    "filters": {"regionName": "PUNE"},
    "limit": 5
})
# → regionID: district_490, regionName: PUNE, parentID: state_27
# 147,423 total regions (states, districts, subdistricts, villages, ULBs)

5. FMD Vaccination Progress by Round

"Show me the latest FMD vaccination round progress in Karnataka"

data = await client.call_tool("4_get_data", {
    "dataset_id": "0055",
    "table_name": "round6",
    "filters": {"district.name": "Bengaluru Urban"},
    "limit": 10
})
# → 588 daily progress records for Round 6
# Sample: date=2024-11-28, cattle_vaccinated=9, buffalo_vaccinated=0

Connecting from Code

import asyncio
from fastmcp import Client

async def main():
    async with Client("http://localhost:8000/mcp") as client:
        # Step 1: Discover datasets
        overview = await client.call_tool("1_know_about_artpark_data", {})
        print(overview.content[0].text)

        # Step 2-4: Follow the workflow...

asyncio.run(main())

Connecting from Claude Desktop

Add to your Claude Desktop MCP config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "artpark": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Then ask Claude: "What are the top dengue-affected districts in Karnataka?"

Connecting from Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "artpark-health": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Deployment

Docker

docker build -t artpark-mcp .
docker run -d -p 8000:8000 --name artpark-server artpark-mcp

Docker Compose (with Jaeger tracing)

docker-compose up -d
# Server: http://localhost:8000/mcp
# Jaeger UI: http://localhost:16686

Architecture

artpark_server.py          # FastMCP server — 4 tools, LLM-optimized docstrings
artpark/
  client.py                # Local data reader — reads CSV + metadata.yaml
observability/
  telemetry.py             # OpenTelemetry middleware (from esankhyiki-mcp)
publicdata/                # Cloned data repo (dsih-artpark/publicdata)
  data/
    0015/                  # Each dataset: CSV files + metadata.yaml
    0034/
    0041/
    ...

Design Principles (inherited from esankhyiki-mcp)

Credits

• Data: ARTPARK Data Science Innovation Hub (DSIH), IISc Bengaluru

• Architecture: Adapted from nso-india/esankhyiki-mcp (MoSPI MCP Server)

• Framework: FastMCP 3.0

• Protocol: Model Context Protocol