ARTPARK Public Data 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

---

Datasets

ID	Name	Category	Rows	Tables
0015	Karnataka Dengue Daily Summary (2017-2024)	Epidemiology	76,114	1
0034	Local Government Directory Region IDs	Geospatial	147,423	1
0041	20th Livestock Census (2019)	Census	30+ per table	4 (+11 state-level)
0055	FMD NADCP Vaccination Progress (Rounds 1-6)	Epidemiology	930-588/round	6
0059	FMD NADCP Vaccination Schedule (Rounds 3-6)	Epidemiology	varies	4
0086	FMD Nationwide Vaccination Data	Epidemiology	varies	1
0087	FMD Nationwide Seromonitoring Data	Epidemiology	238	1
0089	FMD Nationwide Serosurveillance Data	Epidemiology	varies	1

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

Step	Tool	Description
1	1_know_about_artpark_data()	Overview of all 8 datasets. Start here.
2	2_get_tables(dataset_id)	List tables (CSVs) in a dataset
3	3_get_metadata(dataset_id, table_name)	Column schema + filterable values (MANDATORY before step 4)
4	4_get_data(dataset_id, table_name, filters)	Fetch data with optional filters

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:

Investigation	What the AI Did	Key Finding
One Health Hypothesis	Joined dengue + livestock data across 17 districts, computed Spearman correlation	Livestock doesn't predict dengue (r = +0.01)
COVID Damage Forensics	Compared pre/post-COVID immunity for 19 states, ranked crash severity	Average -32.8pp immunity loss nationwide
Needle in a Haystack	Scanned 238 rows for 4 anomaly categories, crowned a winner	Nagaland 2021: vaccination halved immunity
Myth Debunker	Classified all 35 states as GREEN/YELLOW/RED using 2 datasets	Only 4 states (11%) show clear success
FMD Risk Model	Built a 3-component scoring formula, ranked 19 states	AP, Bihar, Gujarat are highest risk

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)

Principle	Implementation
Sequential workflow	Numbered tool prefixes (1_, 2_, 3_, 4_) enforce order
LLM-optimized docstrings	RULES (MUST follow exactly) blocks prevent hallucinated queries
metadata.yaml as source of truth	Column schemas from YAML, not hardcoded
Response guidance	_next_step and _retry_hint keys steer the LLM
Case-insensitive filtering	"Mysuru" matches "MYSURU" in the data

---

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