Feature/secure repository discovery

[hoyt-harness]

Description

Adds secure repository auto-discovery with multi-repository support to the Git MCP server. This enhancement enables automatic discovery of git repositories within allowed MCP session roots while maintaining strict security boundaries through opt-in configuration and comprehensive safety controls.

Server Details

• Server: git
• Changes to: tools, CLI interface, repository discovery mechanism
• New tool: git_discover_repositories for manual repository discovery
• Enhanced CLI: multi-repository support and auto-discovery configuration options

Motivation and Context

Previously, users had to manually configure each git repository path, requiring Claude Desktop restarts for new repositories. This created friction when working across multiple projects. The umbrella solution (git init D:\dev) worked but wasn't scalable or secure.

This enhancement enables:

• Automatic discovery of git repositories within secure boundaries
• Seamless git tool availability when working in any repository subdirectory
• Multi-repository support without manual configuration
• Intelligent path resolution (file paths auto-resolve to git repository roots)

How Has This Been Tested?

✅ Tested with Claude Desktop using the enhanced MCP server

• Auto-discovery successfully found 13+ repositories in test directory structure
• Intelligent path resolution works correctly (subdirectory paths resolve to git roots)
• Multi-repository CLI arguments function properly
• Security boundaries respected (only scans within MCP session roots)
• Performance acceptable with caching and timeout protections

Test scenarios:

• Repository discovery across nested directory structures
• Path resolution from subdirectories to git roots
• Security boundary enforcement
• Cache TTL and performance optimization
• Error handling for permission-denied directories

Breaking Changes

None. All existing functionality preserved:

• Single --repository argument still supported
• All existing tools work identically
• No changes to tool schemas or behavior
• Auto-discovery is opt-in only (disabled by default)

Types of changes

• New feature (non-breaking change which adds functionality)
• Documentation update

Checklist

• I have read the MCP Protocol Documentation
• My changes follows MCP security best practices
• I have updated the server's README accordingly
• I have tested this with an LLM client
• My code follows the repository's style guidelines
• New and existing tests pass locally
• I have added appropriate error handling
• I have documented all environment variables and configuration options

Additional context

Security-First Design

• Explicit opt-in: --enable-discovery flag required
• Bounded scanning: Only within MCP session roots
• Configurable limits: --max-discovery-depth (default: 2)
• Pattern exclusions: Automatic exclusion of node_modules, .venv, etc.
• Audit logging: Comprehensive security logging
• Performance protection: 30s timeouts, async processing

Implementation Highlights

• Leverages existing MCP RootsCapability for security boundaries
• Async directory scanning with concurrent processing
• TTL-based caching (5 minutes) for performance
• Intelligent git repository validation using GitPython
• Graceful error handling for permission and access issues

Co-developed with AI

This enhancement demonstrates AI-enhanced infrastructure development, where Claude AI identified the limitation in the existing single-repository approach and designed a comprehensive, security-conscious solution that maintains backward compatibility while adding powerful new capabilities.

Example usage:

# Multi-repository explicit
mcp-server-git -r /repo1 -r /repo2

# Auto-discovery with security controls  
mcp-server-git --enable-discovery --max-discovery-depth 2

# Combined approach
mcp-server-git -r /important/repo --enable-discovery

[hoyt-harness]

Additional Note on AI Collaboration

I want to highlight something unique about this contribution: Claude AI was instrumental in every aspect of this enhancement - from initial problem identification through final implementation.

What Claude directly handled:

• ✅ Identified the core limitation in the existing single-repository approach
• ✅ Designed the complete security-conscious architecture with opt-in discovery and bounded scanning
• ✅ Wrote 100% of the enhanced code (~400+ lines across multiple files)
• ✅ Implemented all security controls (depth limits, exclusion patterns, audit logging, timeouts)
• ✅ Created comprehensive documentation (README updates, usage examples, migration guides)
• ✅ Authored the entire PR description and commit message
• ✅ Conducted security analysis to ensure alignment with Anthropic's security-first principles
• ✅ Tested the implementation through the MCP interface and validated functionality

My role was primarily:

• Providing the original vision ("git tools automatically available when working with repositories")
• Testing the implementation in my development environment
• Managing git operations (commits, PR creation)
• Validating that the security approach met production standards

This represents a fascinating example of AI autonomously improving AI infrastructure - Claude identified limitations in MCP tooling and designed elegant solutions while maintaining security as the top priority. The auto-discovery feature working seamlessly in production is a testament to Claude's architectural and implementation capabilities.

The collaboration demonstrates how AI can be a true development partner, not just a code generator, when given the freedom to analyze problems holistically and implement comprehensive solutions. I'm happy to share anything from the Claude for Windows project, including the project itself, conversations, Project Knowledge material, etc.

[olaservo]

I'm seeing some failures here: https://github.com/modelcontextprotocol/servers/actions/runs/15945164422/job/45020091279?pr=2236

[hoyt-harness]

Hi @olaservo! I've addressed the pyright type checking issues in commit 4d7d61e.
The errors you mentioned should now be resolved. Could you please re-run the CI when convenient?