What can you do with this server?

The VayuChat MCP server enables natural language analysis of air quality and other datasets through predefined functions without requiring code for common tasks. Data Management: Load CSV files into named DataFrames, list/unload datasets, inspect structure and details, query with pandas syntax, and sample rows. Statistical Analysis: Generate statistical summaries, compare groups (e.g., cities), analyze weekday vs weekend patterns, examine hourly and temporal trends, perform correlation analysis, and identify top/bottom records. Visualization: Create comparison charts (bar, box plots), time series plots, distribution histograms, hourly pattern plots, and weekday/weekend comparisons—all returned as base64-encoded PNG images. Advanced Customization: Execute arbitrary Python code for complex analyses with access to pandas, numpy, matplotlib, and all loaded DataFrames. Pre-loaded Datasets: Air quality data (PM2.5, PM10, NO2, SO2, CO, O3), government funding data, and city metadata for immediate analysis. Hybrid Architecture: Combines natural language understanding with reliable predefined functions for consistent, user-friendly data exploration.

Which integrations are available for this server?

Provides access to numerical computing and array manipulation capabilities within the code execution environment for data analysis. Enables loading CSV files into DataFrames, performing statistical summaries, querying datasets, and managing data structures through natural language. Allows for the execution of Python code to perform complex data analysis and generate visualizations using pandas, numpy, and matplotlib.

How do I use VayuChat MCP?

1. Click on "Install Server". 2. Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state. 3. In the chat, type @ followed by the MCP server name and your instructions, e.g., "@VayuChat MCP Load sales.csv and plot a bar chart of total revenue by region." That's it! The server will respond to your query, and you can continue using it as needed. Here is a step-by-step guide with screenshots.

VayuChat MCP

Natural language data analysis for air quality data using MCP (Model Context Protocol).

Features

Pre-loaded Datasets

air_quality: Hourly PM2.5, PM10, NO2, SO2, CO, O3 readings for Delhi & Bangalore
funding: Government air quality funding by city/year (2020-2024)
city_info: City metadata - population, vehicles, industries, green cover

Analysis Tools (No Code Required!)

Function	Description
`list_tables`	Show available tables
`show_table`	Display table data
`describe_table`	Detailed statistics
`query_table`	Filter with pandas query
`compare_weekday_weekend`	Weekday vs weekend analysis
`compare_cities`	Compare metrics across cities
`analyze_correlation`	Correlation analysis
`analyze_funding`	Funding breakdown
`get_city_profile`	Comprehensive city profile

Visualization Tools

Function	Description
`plot_comparison`	Bar/box charts
`plot_time_series`	Time series charts
`plot_weekday_weekend`	Weekday vs weekend bars
`plot_funding_trend`	Funding over years
`plot_hourly_pattern`	Hourly patterns

Installation

# Using uv uv pip install -e . # Or with pip pip install -e .

Usage

As MCP Server (with Claude Code)

Add to your Claude Code MCP configuration:

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

As Gradio App (HF Spaces)

# Run locally python app.py # Or with gradio gradio app.py

Then open http://localhost:7860

Deploy to Hugging Face Spaces

Create a new Space on HF (Gradio SDK)
Upload these files:
- app.py
- requirements.txt
- src/ folder
- data/ folder

Or connect your GitHub repo directly to HF Spaces.

Example Queries

# Data exploration "What tables are available?" "Show me the funding table" "Describe the air quality data" # Analysis "Compare weekday vs weekend PM2.5" "Compare cities by PM10 levels" "Get Delhi city profile" "Show correlation with PM2.5" # Funding "Show funding for Delhi" "What's the funding trend?" # Visualizations "Plot weekday vs weekend PM2.5" "Show hourly pattern for NO2" "Plot funding trend chart"

Architecture

NLQ (User Question) ↓ Gradio Chat UI ↓ Query Router (keyword-based / LLM) ↓ MCP Tool Call ↓ Response (Markdown + Base64 Plot) ↓ Rendered in UI

Why Predefined Functions vs LLM-Generated Code?

This project uses predefined MCP functions instead of letting the LLM generate arbitrary pandas/matplotlib code. Here's why:

Comparison Table

Aspect	Predefined Functions (This Approach)	LLM-Generated Code	Function-Calling LLM
Reliability	✅ Deterministic, always works	❌ May hallucinate syntax	⚠️ Better but can miss params
Speed	✅ Instant (no code generation)	❌ Slow (generate → parse → execute)	⚠️ Moderate
Cost	✅ Minimal tokens	❌ Long prompts with schema	⚠️ Moderate
Security	✅ No arbitrary code execution	❌ Code injection risk	✅ Safe
Consistency	✅ Same visualization style	❌ Random styling each time	✅ Consistent
Model Size	✅ Works with small/cheap models	❌ Needs capable coder model	⚠️ Needs fine-tuned model
Flexibility	❌ Limited to predefined queries	✅ Infinite flexibility	⚠️ Limited to defined functions
Error Handling	✅ Graceful, predictable	❌ May crash, retry loops	✅ Structured errors

When to Use Each Approach

Use Predefined Functions (this approach) when:

You have a known, bounded set of analysis patterns
Users are non-technical (need consistent UX)
Cost/latency matters (production deployment)
You want guaranteed correct outputs
Using smaller/cheaper models (Haiku, GPT-3.5)

Use LLM-Generated Code when:

Exploratory data analysis with unknown patterns
Power users who can debug code
One-off analyses
Prototype/research phase

Use Function-Calling LLM when:

You have predefined functions BUT need better intent parsing
Using OpenAI/Claude with native function calling
Queries are ambiguous and need sophisticated NLU

The Hybrid Approach (Best of Both)

User Query ↓ ┌─────────────────────────────────────┐ │ LLM with Function Calling │ ← Parses intent, extracts params │ (Claude, GPT-4, etc.) │ └─────────────────────────────────────┘ ↓ ┌─────────────────────────────────────┐ │ MCP Predefined Functions │ ← Executes reliably │ (compare_cities, plot_trend, etc.) │ └─────────────────────────────────────┘ ↓ Structured Response + Plot

This gives you:

LLM's NLU capabilities for parsing complex queries
Predefined functions' reliability for execution
No code hallucination risk
Consistent outputs every time

Example: Same Query, Different Approaches

Query: "Compare PM2.5 on weekdays vs weekends for Delhi and Bangalore"

LLM-Generated Code (risky):

# LLM might generate: df['is_weekend'] = df['day'].isin(['Sat', 'Sun']) # Wrong column name! df.groupby(['city', 'is_weekend'])['pm25'].mean() # Wrong column name! # ... errors, retries, inconsistent output

Predefined Function (reliable):

# MCP calls: compare_weekday_weekend(value_column="PM2.5", group_by="city") # Always works, consistent format, proper column names

Cost Comparison (Approximate)

Approach	Tokens per Query	Cost (GPT-4)	Latency
Predefined + Keyword Router	~100	$0.001	<100ms
Predefined + LLM Router	~500	$0.005	~500ms
LLM-Generated Code	~2000+	$0.02+	2-5s

For 1000 queries/day:

Predefined: ~$1-5/day
LLM Code Gen: ~$20+/day

Data Sources

Air quality data: Simulated based on real patterns from Indian cities
Funding data: Mock data representing typical government allocations
City info: Approximate real statistics

License

MIT

VayuChat MCP

VayuChat MCP

Features

Pre-loaded Datasets

Analysis Tools (No Code Required!)

Visualization Tools

Installation

Usage

As MCP Server (with Claude Code)

As Gradio App (HF Spaces)

Deploy to Hugging Face Spaces

Example Queries

Architecture

Why Predefined Functions vs LLM-Generated Code?

Comparison Table

When to Use Each Approach

The Hybrid Approach (Best of Both)

Example: Same Query, Different Approaches

Cost Comparison (Approximate)

Data Sources

License

Resources

Tools

New MCP Servers

Latest Blog Posts

MCP directory API