claude-maintain

Environment maintenance tool for Claude Code (~/.claude/).

Diagnoses MCP server health, finds sync debris, analyzes tool usage from session logs, and produces a health score with prioritized recommendations.

Install

# With uv (recommended)
uv tool install claude-maintain

# With pip
pip install claude-maintain

# Development (editable)
git clone https://github.com/HermeticOrmus/claude-maintain.git
cd claude-maintain
uv tool install -e .

Commands

maintain health -- MCP Server Health Check

Parses both Claude Code CLI (~/.claude/mcp.json) and Claude Desktop (claude_desktop_config.json) configs. Detects:

• Placeholder API keys (YOUR_OPENAI_API_KEY_HERE)
• Cross-platform path issues (Linux paths on Windows, synced via Syncthing/rsync)
• Duplicate services (e.g., 4 Telegram variants)
• Missing required environment variables

maintain health                  # Show health report
maintain health --generate-clean # Save fixed configs to ~/.claude/reports/

maintain clean -- Environment Cleanup

Finds accumulated debris in ~/.claude/:

• Syncthing conflict files (*.sync-conflict-*)
• Orphaned files (.backup, .broken, .old)
• Stale session logs (configurable age threshold)
• Skill naming inconsistencies (SKILL.md vs skill.md)
• Directory size breakdown

maintain clean                 # Dry run (shows what would be cleaned)
maintain clean --execute       # Move debris to ~/.claude/backups/ (safe, reversible)
maintain clean --max-age 60    # Flag sessions older than 60 days

maintain stats -- Usage Analytics

Stream-parses session JSONL files to extract tool usage patterns:

• Per-tool invocation counts (Read, Bash, Write, MCP tools, etc.)
• Skill usage tracking (which /skills are actually used)
• Cross-reference with filesystem (find never-invoked skills)
• Results cached for fast repeat runs

maintain stats                 # Full scan (all sessions, cached)
maintain stats --recent 20     # Only 20 most recent sessions
maintain stats --no-cache      # Force fresh parse

maintain optimize -- Health Score + Recommendations

Runs all three modules and synthesizes a weighted health score (0-100):

Produces prioritized recommendations (CRITICAL / HIGH / MEDIUM / LOW) and saves a markdown report to ~/.claude/reports/.

maintain optimize              # Full analysis + score

JSON Output

All commands support --json for machine-readable output:

maintain --json health         # JSON health report
maintain --json stats          # JSON usage data
maintain --json optimize       # JSON score + recommendations

Safety

• Dry run by default -- maintain clean shows what it would do without acting
• Backups before deletion -- --execute moves files to ~/.claude/backups/maintain-{timestamp}/
• Never overwrites configs -- --generate-clean writes fixed versions to ~/.claude/reports/
• Never logs secrets -- API tokens are flagged as a security concern but never displayed in reports

Who is this for?

Anyone who:

• Uses Claude Code across multiple machines (Syncthing, rsync, cloud sync)
• Has accumulated MCP servers they've experimented with
• Wants to know which of their 100+ skills are actually used
• Has never cleaned ~/.claude/ and suspects it's grown large

Requirements

• Python 3.10+
• Claude Code installed (~/.claude/ directory exists)

License

MIT