Summary

This PR introduces a data-driven default toolset configuration to improve performance, reduce inference costs, and provide a better out-of-the-box experience for GitHub MCP Server users.

Problem

The current "all tools enabled by default" approach causes several issues:

• Excessive context: 102 tools consuming 64.6k tokens by default
• Tool selection degradation: LLMs struggle with too many tool options
• Higher inference costs: Large context windows increase computational expenses
• User feedback: Most users don't customize toolsets and experience performance issues

Changes

New Default Toolsets (Data-Driven Selection)

Based on real-world usage analysis across all sessions, the new default includes the 5 most-used toolset groups:

• context - Current user and GitHub context
• repos - Repository operations
• issues - Issue tracking
• pull_requests - PR workflows
• users - User information

These cover almost all top 20 most frequently used tools.

Code Changes

1. pkg/github/tools.go:

   • AddedToolsetMetadata struct to hold toolset ID and description
   • Created metadata constants for all toolsets
   • ImplementedGetDefaultToolsetIDs() to return curated default list
   • AddedGenerateToolsetsHelp() for improved CLI help text
   • Moved stargazer tools fromrepos to new dedicatedstargazers toolset

2. cmd/github-mcp-server/main.go:

   • Updated default flag behavior to useGetDefaultToolsetIDs() when no toolsets specified
   • Improved help text withGenerateToolsetsHelp()

3. README.md:

   • Added prominent default toolset configuration section
   • Reorganized toolset documentation for better discoverability
   • Addedstargazers toolset to available toolsets table

4. docs/remote-server.md:

   • Addedstargazers toolset to remote server documentation

Performance Improvements compared to old default("all")

Token Reduction

• Before: 102 tools, 64.6k tokens
• After: 52 tools, 30.3k tokens
• Savings:53% reduction in token usage

User Experience Benefits

• ✅ Faster LLM responses with smaller context windows
• ✅ More accurate tool selection from focused set
• ✅ Lower inference costs
• ✅ Better out-of-the-box behavior for common workflows

Future Work

• Split large toolsets (repos,pull_requests) into more specialized subgroups

Testing

• Verified default toolsets load correctly when no--toolsets flag provided
• Confirmed--toolsets all still enables all toolsets
• Validated custom toolset combinations work as expected
• Checked README and documentation updates render correctly