CERN GitLab MCP Server

Features

• 14 MCP tools for searching, browsing, and analyzing CERN GitLab repositories

• Dual-mode operation — stdio (single-user) and HTTP (multi-user) modes

• CLI tool (cerngitlab-cli) for direct command-line usage

• Public access — works without authentication for public repositories

• Multi-user support — HTTP mode enables centralized deployment for multiple users

• HEP-focused — dependency parsing for Python and C++ ecosystems, binary detection for .root files, issue search

• Robust — rate limiting, retries with exponential backoff, graceful error handling

Installation

Requires Python 3.10+.

Quickstart (recommended)

No installation needed — just use uvx to run directly:

uvx cerngitlab-mcp

From PyPI

pip install cerngitlab-mcp

From source

git clone https://github.com/MohamedElashri/cerngitlab-mcp
cd cerngitlab-mcp
uv sync

Configuration

All settings are configured via environment variables prefixed with CERNGITLAB_:

Authentication

The server works in two modes:

• Without token — Access to all public repositories. Sufficient for most HEP code discovery.

• With token — Additional access to internal/private projects, code search, and wiki pages.

To create a token:

1. Go to https://gitlab.cern.ch/-/user_settings/personal_access_tokens

2. Create a token with read_api scope

3. Set CERNGITLAB_TOKEN=glpat-xxxxxxxxxxxx

Note: The code search (search_code), issue search (search_issues), and wiki (get_wiki_pages) tools require authentication on CERN GitLab.

Usage

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "cerngitlab": {
      "command": "uvx",
      "args": ["cerngitlab-mcp"],
      "env": {
        "CERNGITLAB_TOKEN": "glpat-xxxxxxxxxxxx"
      }
    }
  }
}

Note for macOS users: If you see an error about uvx not being found, you may need to provide the absolute path. Claude Desktop does not support ~ or $HOME expansion.

1. Run which uvx in your terminal to find the path (e.g., /Users/yourusername/.local/bin/uvx).

2. Use that absolute path in the command field:

"command": "/Users/yourusername/.local/bin/uvx"

Claude Code

Project-specific (default) — installs in the current directory's configuration:

claude mcp add cerngitlab-mcp -- uvx cerngitlab-mcp

Global — installs for your user account (works in all projects):

claude mcp add --scope user cerngitlab-mcp -- uvx cerngitlab-mcp

To include authentication, add -e CERNGITLAB_TOKEN=glpat-xxxxxxxxxxxx before the --:

# Example: Global installation with token
claude mcp add --scope user -e CERNGITLAB_TOKEN=glpat-xxxxxxxxxxxx cerngitlab-mcp -- uvx cerngitlab-mcp

Manual Configuration — you can also manually edit your global config at ~/.claude.json (on Linux/macOS) or %APPDATA%\Claude\claude.json (on Windows):

{
  "mcpServers": {
    "cerngitlab": {
      "command": "uvx",
      "args": ["cerngitlab-mcp"],
      "env": {
        "CERNGITLAB_TOKEN": "glpat-xxxxxxxxxxxx"
      }
    }
  }
}

GitHub Copilot

Add to your VS Code settings.json:

{
  "mcp": {
    "servers": {
      "cerngitlab": {
        "command": "uvx",
        "args": ["cerngitlab-mcp"],
        "env": {
          "CERNGITLAB_TOKEN": "glpat-xxxxxxxxxxxx"
        }
      }
    }
  }
}

Or add a .vscode/mcp.json to your project:

{
  "servers": {
    "cerngitlab": {
      "command": "uvx",
      "args": ["cerngitlab-mcp"],
      "env": {
        "CERNGITLAB_TOKEN": "glpat-xxxxxxxxxxxx"
      }
    }
  }
}

Gemini CLI

Add to your ~/.gemini/settings.json:

{
  "mcpServers": {
    "cerngitlab": {
      "command": "uvx",
      "args": ["cerngitlab-mcp"],
      "env": {
        "CERNGITLAB_TOKEN": "glpat-xxxxxxxxxxxx"
      }
    }
  }
}

Direct usage

Stdio Mode (Single-User, Default)

# Run with uvx (no install needed)
uvx cerngitlab-mcp

# Or if installed from PyPI
cerngitlab-mcp

# Explicit stdio mode
cerngitlab-mcp --mode stdio

# Or from source
uv run cerngitlab-mcp

# With authentication
CERNGITLAB_TOKEN=glpat-xxx uvx cerngitlab-mcp

HTTP Mode (Multi-User)

# HTTP mode for multi-user deployment
cerngitlab-mcp --mode http --host 0.0.0.0 --port 8080

# Or use environment variables
CERNGITLAB_HTTP_MODE=true CERNGITLAB_HOST=0.0.0.0 CERNGITLAB_PORT=8080 cerngitlab-mcp

# Dedicated HTTP entry point
cerngitlab-mcp-http

Mode Selection

• --mode stdio - Single-user mode using stdin/stdout (default)

• --mode http - Multi-user mode using HTTP API

• --mode auto - Auto-detect based on environment variables

The server automatically detects HTTP mode if CERNGITLAB_HTTP_MODE, CERNGITLAB_HOST, or CERNGITLAB_PORT environment variables are set.

Tools

For detailed parameter documentation, see docs/dev.md.

Example Prompts

Search for repositories

"Search CERN GitLab for Python repositories related to ROOT analysis and show me the most starred ones"

Understand a project

"Get the README and file structure of the lhcb/DaVinci project on CERN GitLab"

Find fitting examples

"Search for repositories on CERN GitLab that use RooFit and show me example fitting code"

View LHCb software stack code

"Search the LHCb sim11 stack for the initialization routines in the Boole project"

Analyze a project structure

"Inspect the lhcb/allen project to understand its build system, dependencies, and CI pipeline configuration"

Find usage context

"Search for issues related to 'segmentation fault' in the atlas/athena project to see if others have encountered this"

Track releases

"List the recent releases of lhcb/DaVinci and show me the release notes for the latest version"

Find framework configurations

"Search for Gaudi framework configuration files on CERN GitLab and show me examples"

Development

See docs/dev.md for development setup, project structure, testing, and release instructions.

Multi-User HTTP Deployment

For centralized deployments serving multiple users, use HTTP mode:

Setup

# Start HTTP server
cerngitlab-mcp --mode http --host 0.0.0.0 --port 8080

# Or with environment variables
CERNGITLAB_HTTP_MODE=true CERNGITLAB_HOST=0.0.0.0 CERNGITLAB_PORT=8080 cerngitlab-mcp

User Authentication

For demo purposes, users can be configured via environment variables:

# Set up demo users
export CERNGITLAB_DEMO_USER_alice=glpat-alice-token
export CERNGITLAB_DEMO_USER_bob=glpat-bob-token

Users authenticate with Authorization: Bearer demo-alice header.

API Endpoints

• GET / - Server information

• GET /health - Health check

• GET /tools - List available tools (authenticated)

• POST /tools/{tool_name} - Execute specific tool (authenticated)

• POST /mcp - Generic MCP endpoint (authenticated)

Example Usage

# List available tools
curl -H "Authorization: Bearer demo-alice" http://localhost:8080/tools

# Execute a tool
curl -X POST -H "Authorization: Bearer demo-alice" \
     -H "Content-Type: application/json" \
     -d '{"arguments": {"query": "ROOT"}}' \
     http://localhost:8080/tools/search_projects

CLI Tool

A command-line interface is also available for direct usage without the MCP server:

# Install or use with uvx
uvx cerngitlab-cli

# Test connectivity
cerngitlab-cli test-connection

# Search for projects
cerngitlab-cli search-projects --query "ROOT analysis" --language python

# Get project info
cerngitlab-cli get-project-info --project lhcb/DaVinci

# Search code
cerngitlab-cli search-code --search-term "RooFit" --per-page 10

# Inspect project structure
cerngitlab-cli inspect-project --project lhcb/allen

All commands output JSON to stdout for easy piping and composition. See cerngitlab-cli --help for the full list of commands.

Skill File

A detailed skill file (SKILL.md) is available with:

• Complete documentation of all 14 tools

• Input/output specifications

• Usage examples

• Authentication requirements

This can be used by LLMs or agents to understand the available tools and how to use them.

License

AGPL-3.0