An MCP server and CLI tool that allow LLMs to interact with CERN ROOT files.
Repository Mirror: This repository is mirrored to CERN GitLab for CERN users. PyPI releases are published via GitHub Actions using attestation-based trusted publishing.
ROOT-MCP empowers Large Language Models (LLMs) to natively understand and analyze CERN ROOT files.
By exposing a set of specialized tools via the Model Context Protocol (MCP) or a token-efficient CLI interface, it turns Claude (and other MCP-compliant agents) into capable physics research assistants that can:
- Inspect ROOT file structures (Trees, RNTuples, Branches, Histograms)
- Analyze data distributions (Compute Histograms, Statistics)
- Compute kinematic quantities (Invariant Mass)
- Visualize results (Plot 1D/2D histograms directly)
- Filter data using physics cuts ("selections")
Why this matters: Instead of asking an LLM to "write a script" that you have to debug and run, you can ask the LLM to "Check the muon pT distribution in this file" and it will just do it.
ROOT-MCP provides two ways to interact with ROOT files:
- Full JSON-RPC protocol support
- Structured input/output for programmatic use
- Best for: MCP-compliant LLM clients, automated workflows
- Human-readable output by default
- Simpler architecture (no server process)
- Best for: Direct LLM interaction, debugging, scripting
Both interfaces share the same analysis backend. MCP also provides resource discovery and server capability information.
pip install root-mcpOptional: For remote file access via XRootD protocol:
pip install "root-mcp[xrootd]"Fastest path — no config file needed:
# MCP Server
root-mcp --data-path /path/to/your/data
# CLI (token-efficient)
root-cli -d /path/to/your/data lsOr set an environment variable once:
export ROOT_MCP_DATA_PATH=/path/to/your/data# Codex
codex mcp add root-mcp -- uvx root-mcp --data-path /absolute/path/to/data
# Claude Code (project scope)
claude mcp add --transport stdio --scope project root-mcp -- uvx root-mcp --data-path /absolute/path/to/dataUse /mcp in your client to check the connection. See the
LLM integration guide for development checkouts,
Antigravity CLI, Qwen Code, Pi/OMP, VS Code, Cursor, Zed, community harnesses,
authenticated HTTP and bounded analysis workflows.
The server uses MCP Python SDK 2.3+ and the 2026-07-28 protocol, with SDK compatibility for older clients. It provides discovery instructions, explicit tool safety hints, validated inputs, structured results and readable resource catalogs. Remote deployments retain authentication, resource ACLs and quotas. See the v0.3.1 release notes for the full change list.
The CLI provides a token-efficient, human-readable interface that provides significant token savings compared to the MCP JSON protocol.
# List files
root-cli ls
# Inspect a file
root-cli inspect /data/sample.root
# Create histogram with fit
root-cli histogram /data/sample.root events muon_pt --bins 100 --fit gaussian
# Read data with selection
root-cli read /data/sample.root events met muon_pt --selection "met > 50"
# Plot results
root-cli plot1d /tmp/root_mcp/muon_pt_hist.json -o plot.png --title "Muon pT"Ask your LLM: "Plot the muon pT distribution"
The LLM generates:
root-cli histogram /data/sample.root events muon_pt --bins 50 && \
root-cli plot1d /tmp/root_mcp/muon_pt_hist.json -o muon_pt.png --title "Muon pT Distribution"See docs/skills/root-cli.md for complete command reference with examples.
# Core mode (lightweight, no scipy/matplotlib needed)
root-mcp --data-path /data --mode core
# Explicit stdio transport (same runtime as the compatibility default)
root-mcp serve-stdio --data-path /data
# Extended mode with native ROOT, restricted to one directory
root-mcp --data-path /data --enable-root --allowed-root /data
# Remote XRootD resource, no YAML needed
root-mcp --resource cms=root://xrootd.cern.ch//store --allow-remote --mode extended
# Docker / container — fully env-var driven
ROOT_MCP_DATA_PATH=/data ROOT_MCP_MODE=extended ROOT_MCP_EXPORT_PATH=/exports root-mcp
# Quiet server (only warnings+) with a cache increase
root-mcp --data-path /data --log-level WARNING --cache-size 100root-mcp serve-http serves the MCP Streamable HTTP endpoint for central
deployments. It requires explicit auth, Origin validation, restrictive central
policy, and safe bind-address settings before it starts.
For shared deployments, start with the operator docs:
docs/operator/central_deployment.md
and docs/operator/security_checklist.md.
Restrictive starter configs and Kubernetes manifests live in
examples/central/.
Operators can preview and apply export retention with
root-mcp cleanup-exports --config /etc/root-mcp/config.yaml --dry-run.
The external HTTP smoke script at
scripts/smoke_external_http_client.py
starts a temporary central server and checks it with the MCP Streamable HTTP
client.
Generate a starter config (optional):
root-mcp init --permissive # creates config.yaml pre-filled with current directoryManual config file — for persistent settings, remote resources, or native ROOT:
server:
mode: "extended" # "core" or "extended"
resources:
- name: "my_analysis"
uri: "file:///path/to/data"
allowed_patterns: ["*.root"]
security:
allowed_roots: [] # empty = any local path is accessible (permissive)Local-use warning: The permissive
allowed_roots: []default is intended for trusted local stdio sessions. Do not expose that configuration as a shared HTTP service; restrict roots and require an authenticated deployment profile before central or multi-user use.
Mode Selection:
mode: "core"— Lightweight: file operations and basic statisticsmode: "extended"— Full analysis: histograms, fitting, kinematics, correlations
Switch modes at runtime with the switch_mode tool — no restart required.
Add to your claude_desktop_config.json:
{
"mcpServers": {
"root-mcp": {
"command": "root-mcp",
"args": ["--data-path", "/path/to/your/data"]
}
}
}Or with a persistent config file:
{
"mcpServers": {
"root-mcp": {
"command": "root-mcp",
"env": {
"ROOT_MCP_CONFIG": "/path/to/config.yaml"
}
}
}
}ROOT-MCP organizes analysis capabilities into analysis tiers:
- Core Mode: File I/O, data reading, and basic statistics
- Extended Mode: Full analysis capabilities including fitting, kinematics, and correlations
The mode is controlled via configuration, and the server automatically loads only the components you need. Runtime mode switching is also available.
ROOT-MCP can optionally integrate with a native ROOT/PyROOT installation to unlock capabilities beyond what uproot provides:
run_root_code: Execute arbitrary PyROOT/Python code and get structured resultsrun_rdataframe: Compute histograms using ROOT's RDataFrame (no boilerplate needed)run_root_macro: Execute C++ ROOT macros viagROOT.ProcessLine
This feature is entirely optional — ROOT-MCP works fully without ROOT installed. When ROOT is available and enabled, these additional tools appear automatically.
Requirements: A working ROOT installation (via conda-forge, system package, or binary tarball). ROOT is not pip-installable at this time.
Enable it by setting enable_root: true in your config.yaml:
features:
enable_root: true
# Optional: tune execution settings
root_native:
execution_timeout: 60
working_directory: "/tmp/root_mcp_native"Use get_server_info to check ROOT availability at runtime:
{
"root_native_available": true,
"root_native_enabled": true,
"root_version": "6.32/02",
"root_features": {"rdataframe": true, "roofit": true, "tmva": false}
}The full documentation site is built with Sphinx and covers installation, configuration, all 20 MCP tools, LLM integration patterns, and the developer guide with auto-generated API reference.
Read online: The docs are hosted at root-mcp docs
pip install "root-mcp[docs]"
./scripts/build_docs.sh
# open docs/_build/html/index.htmlFor live-reload while writing docs:
cd docs && make livehtmlHighlights:
- User Guide — installation, quickstart, modes, configuration, LLM integration
- Tool Reference — complete catalogue of all tools and their JSON payloads
- CLI Reference — complete command reference for root-cli with examples
- Developer Guide — architecture, module overview, dev setup, contributing
- API Reference — auto-generated from source docstrings
If you use ROOT-MCP in your research, please cite:
@software{root_mcp,
title = {ROOT-MCP: Production-Grade MCP Server for CERN ROOT Files},
author = {Mohamed Elashri},
year = {2025},
url = {https://github.com/MohamedElashri/root-mcp}
}MIT License - see LICENSE for details.