rostaing-mcp 0.1.1

A universal Model Content Protocol (MCP) for Pandas and Polars DataFrames, designed to be used as a tool by Large Language Models (LLMs).

Rostaing's Model Content Protocol (rostaing-mcp)

A universal Model Content Protocol (MCP) for Pandas and Polars DataFrames, designed to be used as a powerful tool by Large Language Models (LLMs) like GPT-4, Claude, and Mistral.

This package allows LLMs to interact securely with data, performing advanced tasks ranging from simple filtering to statistical hypothesis testing and data visualization (returning viewable images).

Key Features

• Library Agnostic: Seamlessly handles pandas and polars DataFrames.
• Visual Intelligence: Generates Plotly charts returning Base64 PNG images, allowing LLMs to "see" and generate charts.
• Statistical Suite: Built-in support for T-tests, ANOVA, Chi-Squared, Normality tests (Shapiro), and Survival Analysis (Logrank).
• Smart NLU (Fuzzy Matching): Automatically corrects typos in column names (e.g., understands that "salary" refers to "Salary_USD").
• Stateful Analysis: Filters and modifications persist within the agent's session context.

Installation

pip install rostaing-mcp

Note: This will also install necessary dependencies like plotly, kaleido (for image generation), pingouin, and scipy.

Quick Start

import pandas as pd
from rostaing_mcp import DataFrameAgent, DataFrameToolHandler

# 1. Create a sample DataFrame
data = {
    'employee_name': ['Rostaing', 'Lucrèce', 'Isnard', 'Charline', 'Dacier', 'Nora'],
    'salary': [100000, 70000, 85000, 60000, 95000, 62000],
    'experience_level': ['Expert', 'Senior', 'Manager', 'Junior', 'Principal', 'Mid-Level'],
    'department': ['AI', 'Sales', 'AI', 'Sales', 'AI', 'Sales']
}
df = pd.DataFrame(data)

# 2. Initialize the core agent (Works with Polars too!)
data_agent = DataFrameAgent(df, source_description="Employee salary data")

# 3. Wrap it in the tool handler
df_tool = DataFrameToolHandler(data_agent)

# --- EXAMPLES OF DIRECT USAGE ---

# A. Inspect Data
print(df_tool.get_schema())

# B. Statistical Test (e.g., T-test between groups)
# Note: Handles fuzzy matching if you type 'Department' instead of 'department'
print(df_tool.perform_t_test(a='salary', group='department')) 

# C. Visualization (Returns Base64 Image string)
# The LLM can call this to generate a chart
image_data = df_tool.plot_bar_chart(x='employee_name', y='salary', color='department')
print("Chart generated successfully (Base64 data ready).")

Integration with LLM Agents (e.g., Upsonic, LangChain)

To let an LLM use all available tools, you can pass the methods dynamically or wrap them in a proxy class.

from upsonic import Agent, Task

# Get the list of all callable tools for the LLM
tools_list = df_tool.get_all_tools() 

task = Task(
    description="Analyze the salary distribution and plot a bar chart by department.",
    tools=tools_list
)

agent = Agent(model="openai/gpt-4o", name="Data Analyst")
result = agent.do(task)
print(result)

Available Tools

📊 Visualization

• plot_histogram, plot_bar_chart, plot_line_chart
• plot_scatter_plot, plot_box_plot, plot_violin_plot
• plot_heatmap, plot_pie_chart, plot_3d_scatter, and more.

🧮 Statistics

• get_summary_statistics, get_correlation_matrix
• perform_normality_test (Shapiro-Wilk)
• perform_t_test (Student's t-test)
• perform_anova (One-way ANOVA)
• perform_chi2_test (Independence)
• perform_logrank_test (Survival analysis)

🛠 Manipulation

• filter_rows (Complex conditions supported)
• sort_values
• select_columns

Useful Links

• Author's LinkedIn
• Author's YouTube Channel
• GitHub Repository
• PyPI Project Page