ClaudeKit

claudekit

Smart guardrails and workflow automation for Claude Code - catch errors in real-time, save checkpoints, and enhance AI coding with expert subagents

🚀 Installation

⚠️ Requires: Claude Code Max plan (for optimal token usage) • Node.js 20+

npm install -g claudekit
# or: yarn global add claudekit
# or: pnpm add -g claudekit

⚡ Quick Start

# Initialize in your project
claudekit setup

# In Claude Code, try these commands:
/git:status                      # Groups changes by type & suggests commit strategy
/spec:create "your next feature" # Researches codebase & writes full spec
/code-review                     # 6 specialized agents analyze code in parallel

🎯 What It Does

Claudekit acts as your safety net while coding with Claude:

Before: Claude adds 'any' type → ❌ Lost type safety discovered in code review
After:  Claude adds 'any' type → ✅ Instant block: "Use specific type: User | null"

Before: Risky refactor fails → ❌ Git archaeology to find working version  
After:  Risky refactor fails → ✅ One command: /checkpoint:restore

Before: Claude breaks tests → ❌ You discover it after Claude finishes
After:  Claude breaks tests → ✅ Claude sees error immediately and fixes it

Before: Ask Claude to review → ❌ Shallow analysis, general feedback, sequential execution
After:  Use /code-review → ✅ 6 specialized agents analyze in parallel, dynamically pull technology-specific expertise

Key Features

🗺️ Instant Codebase Navigation

Claude sees your entire project structure automatically - skips most discovery searches.

• Navigate directly to code - Jump straight to files, functions, and classes
• No trial-and-error - Avoid sifting through test files, docs, and irrelevant matches
• Confident code access - Claude knows exactly what exists and where
• See relationships - Claude understands dependencies and architecture instantly
• Automatic setup - Runs invisibly on first prompt, updates as you code

Codebase map guide →

🔍 Comprehensive Code Review

• Multi-aspect analysis: 6 parallel agents for architecture, security, performance, testing, quality, and documentation
• Smart targeting: Automatically selects relevant review agents based on file types and scope
• Structured reports: Prioritized issues (Critical/High/Medium) with type icons and actionable solutions
• Quality metrics: Scoring system and issue distribution tracking for measurable improvements

🚀 Iterative Spec Implementation

• 6-phase workflow: Implementation → Test Writing → Code Review → Iterative Improvement → Commit → Progress Tracking
• Dynamic agent selection: Real-time detection of available specialized agents with automatic fallbacks
• Quality gates: Each phase includes validation before proceeding to ensure high-quality deliverables
• Atomic commits: Standardized commit messages with comprehensive change documentation

Spec workflow guide →

🛡️ Real-time Error Prevention

• 🔒 Advanced Bash Security Analysis: Comprehensive protection against sensitive file access via shell commands
  • Intelligent Command Parsing: Analyzes complex bash pipelines, variable assignments, and shell constructs
  • Pattern Detection: Identifies sensitive file access via xargs, find, pipes, and command substitution
  • Path Extraction: Safely processes globs, variables, and @file references in curl/httpie uploads
  • Security Heuristics: Detects risky patterns like find -name '*.env' | xargs cat before execution
• 🔒 Enhanced Sensitive File Protection: Comprehensive security with 195+ patterns across 12 categories including cloud credentials, cryptocurrencies, databases, tokens, and production data
• Multi-category Protection: Environment files, SSH keys, cloud provider credentials (AWS/Azure/GCP), package manager auth, crypto wallets, production databases
• TypeScript Guard: Blocks any types and type errors as Claude edits
• Linting: Catches style issues immediately
• Anti-patterns: Prevents code replacement with comments
• Test Runner: Runs relevant tests on file changes
• Multi-tool ignore file support: Respects .agentignore, .aiignore, .aiexclude, .geminiignore, .codeiumignore, .cursorignore

💾 Git Checkpoint System

• Auto-save: Creates checkpoints when Claude stops
• Easy restore: /checkpoint:restore to undo changes
• Checkpoint management: List, restore, or clean up checkpoints

Checkpoint guide →

🤖 AI Subagents

• Research Expert: research-expert for comprehensive parallel research with 90% time reduction, automatic source analysis, and structured reporting
• Code Search: code-search for fast, parallel codebase searches with 3-10x speed improvements
• Triage Expert: triage-expert for initial problem diagnosis and routing to specialists
• Domain experts: TypeScript, React, database, testing, infrastructure, documentation specialists
• Code review: Dedicated code-review-expert agent with 6 focused analysis aspects
• Framework specialists: Next.js, Nest.js, AI SDK, and build tool experts
• Performance optimization: Intelligent hook management with disableHooks configuration to skip unnecessary validations during specialized operations
• Mandatory delegation: All technical issues MUST be handled by specialized subagents
• Proactive help: Run /agents-md:init to update AGENTS.md/CLAUDE.md for automatic subagent delegation

📊 Hook Performance Profiling

• Performance monitoring: Analyze hook execution time and output size with claudekit-hooks profile
• Batch analysis: Automatically profiles all configured hooks from .claude/settings.json
• Statistical analysis: Support for multiple iterations with averaging for accurate measurements
• Smart warnings: Color-coded alerts for slow hooks (>5s) and output size limits
• Claude Code integration: Specialized limits for UserPromptSubmit hooks that inject context
• Token estimation: Calculate estimated token usage for hook outputs

Hook profiling guide →

🧠 Enhanced AI Reasoning

• Automatic enhancement: The thinking-level hook invisibly enhances Claude's reasoning without manual prompt engineering
• 4 configurable levels: From disabled (0) to maximum reasoning (3) with smart defaults
• Invisible operation: Keywords are injected silently without user awareness
• Lightweight: < 5ms overhead with minimal token usage
• Session-aware: Activates on first user prompt to enhance AI reasoning throughout the session

Thinking level guide →

🎛️ Session-Based Hook Control

• Temporary control: Disable/enable hooks for individual Claude Code sessions
• Session isolation: Changes only affect your current session, not other developers
• Smart state management: 4-state system (enabled/disabled/not-configured/not-found)
• Fuzzy matching: Partial hook names with intelligent suggestions
• Zero permanent changes: No configuration files modified

Session hook control guide →

🔬 Multi-Agent Research System

• Parallel Information Gathering: Launch 5-10 research agents simultaneously for 90% faster investigation
• Intelligent Query Classification: Automatically determines research strategy (breadth-first, depth-first, or simple factual)
• Comprehensive Source Analysis: Cross-references findings across multiple domains with automatic citation management
• Structured Research Reports: Executive summaries, key findings, detailed analysis, and methodology documentation
• Filesystem Artifacts: Subagents write full reports to /tmp/research_*.md for efficient synthesis and token optimization

📝 Smart Commands

• /research [query] - Deep parallel research with specialized subagents and structured reporting
• /code-review - Multi-aspect code reviews with 6 parallel agents
• /git:commit - Creates commits following your project's conventions
• /git:checkout - Smart branch creation and switching with conventional naming
• /git:ignore-init - Initialize AI-safe .gitignore patterns for sensitive file protection
• /validate-and-fix - Runs all quality checks and fixes issues
• /spec:create - Generates comprehensive specifications
• /spec:execute - Implements specifications with iterative quality assurance workflow
• /create-subagent - Build custom AI assistants
• /hook:disable [hook-name] - Temporarily disable hooks for current session
• /hook:enable [hook-name] - Re-enable disabled hooks for current session
• /hook:status [hook-name] - Show hook status (enabled/disabled/not-configured)

CLI Commands

claudekit setup              # Interactive setup wizard
claudekit setup --yes        # Quick setup with defaults
claudekit setup --all        # Install everything (all agents)
claudekit setup --user       # Install in user directory (~/.claude) instead of project
claudekit list               # Show all components (hooks, commands, agents, config)
claudekit list agents        # List available agents with token counts
claudekit list commands      # List available commands with token counts
claudekit show agent <id>    # Display agent prompt for external use
claudekit show command <id>  # Display command prompt for external use
claudekit show agent <id> -f json    # Output agent metadata and content as JSON
claudekit show command <id> -f json  # Output command metadata and content as JSON
claudekit doctor             # Check your installation and validate documentation links

# Tool status checking
claudekit status stm         # Check Simple Task Master (STM) installation and initialization status

# Linting and validation
claudekit lint-commands      # Lint slash command markdown files
claudekit lint-agents        # Lint agent markdown files (formerly lint-subagents)

# Hook management and profiling
claudekit-hooks run <hook>   # Execute a specific hook manually
claudekit-hooks profile      # Profile all configured hooks performance
claudekit-hooks profile <hook> --iterations 5  # Profile specific hook with averaging

# Session-based hook control (within Claude Code sessions)
claudekit-hooks disable <hook>  # Disable hook for current session
claudekit-hooks enable <hook>   # Re-enable hook for current session  
claudekit-hooks status <hook>   # Show hook status (enabled/disabled)

Using Prompts with External LLMs

The show command extracts agent and command prompts for use with Claude Code's non-interactive mode and other AI coding tools:

# Extract agent prompt for use with Claude Code CLI
EXPERT=$(claudekit show agent typescript-expert)
cat src/app.ts | claude -p --append-system-prompt "$EXPERT" "Review this code"

# Get structured metadata for integration with tools
claudekit show agent typescript-expert --format json | jq '.description'

# Works with any AI coding CLI tool
claudekit show agent react-expert | amp -x "Fix React hooks violations"

See the [external LLMs guid

…