MCP Refana Server

Model Context Protocol (MCP) server for Mu2e refana analysis workflows. Exposes core analysis functions as MCP tools for integration with agentic frameworks.

Environment Setup

This project requires the Mu2e analysis environment. On Mu2e gpvm's, activate the environment first:

source /cvmfs/mu2e.opensciencegrid.org/env/ana/current/bin/activate

This provides:

• pyutils - Python interface to EventNtuple

• uproot - ROOT file handling

• awkward - Ragged array support

• numpy, scipy, scikit-learn, xgboost, matplotlib

Related MCP server: rucio-mcp

Features

• Event Counting: Signal/background extraction in kinematic regions

• Cut Analysis: Definition and application of analysis cuts

• Sensitivity Computation: Discovery significance, CLs limits, Feldman-Cousins intervals

• ML Selection: BDT-based signal/background discrimination setup

• Analysis Summaries: Dataset-level analysis results

Installation

# Activate Mu2e analysis environment
source /cvmfs/mu2e.opensciencegrid.org/env/ana/current/bin/activate

# Install mcp-refana in development mode
cd mcp-refana
pip install -e .

Running the Server

Important: Always activate the Mu2e environment before running.

# Source Mu2e environment
source /cvmfs/mu2e.opensciencegrid.org/env/ana/current/bin/activate

# Start server with stdio transport (default)
mcp-refana-server --transport stdio

Or with HTTP transport:

source /cvmfs/mu2e.opensciencegrid.org/env/ana/current/bin/activate
mcp-refana-server --transport streamable-http

Available Tools

healthcheck()

Returns server status.

count_signal_background(mom_list, time_list, mc_list, mom_range, time_range, sign)

Extract signal and background event counts in a defined kinematic window.

Parameters:

• mom_list: Reconstructed momentum values

• time_list: Reconstructed time values

• mc_list: MC truth labels

• mom_range: [min, max] momentum window

• time_range: [min, max] time window

• sign: "minus" (electrons) or "plus" (positrons)

Returns: Signal/background counts and statistics

analyze_cuts(cut_definitions, sign)

Initialize analysis cut configuration.

compute_sensitivity(s_expected, b_expected, b_sys_uncertainty, exposure, efficiency)

Compute physics sensitivity metrics:

• Asimov significance Z

• 95% CLs upper limit

• 90% Feldman-Cousins confidence interval

initialize_ml_selector()

Set up machine learning event selector.

get_cut_efficiency(cuts_applied, n_total, n_surviving)

Calculate efficiency metrics for applied cuts.

summarize_analysis(dataset_name, sign, n_events_processed, signal_count, background_count, ml_score_mean)

Generate analysis summary for a dataset.

Integration with HEP-Multiagent

Register this server in the multiagent's config.py:

"refana": {
    "command": ["mcp-refana-server", "--transport", "stdio"],
    "description": "Mu2e refana analysis server",
}

Then use in worker nodes for agentic analysis workflows.

Architecture

src/mcp_refana/
├── __init__.py
├── server.py              # FastMCP server setup
└── mcp_tools/
    ├── __init__.py
    └── refana_wrappers.py # Tool definitions wrapping refana code

The server wraps the core refana analysis classes (Count, Analyze, MLSelect, SensitivityAnalyzer) as MCP tools, making them accessible to agents via tool calls.

Dependencies

• mcp: Model Context Protocol framework

• refana: Core analysis code (count, analyze, ML selection, sensitivity)

• numpy, scipy: Numerical computing

• scikit-learn, xgboost: Machine learning

• uproot, awkward: ROOT file handling

• matplotlib: Visualization

Development

Run tests:

pytest tests/

Format code:

ruff check . --fix

Contact

Sophie Middleton, Huma Jafree