Ag-Kit

Overview

ag-kit is an "antigravity-first" agent engineering kit — a framework for building AI agent projects with built-in safety, rules, skills, specialist agents, workflows, persistent memory, and MCP guidance. At 7,800+ GitHub stars, it has rapidly established itself as a security-conscious alternative to less-guarded agent frameworks.

The term "antigravity" refers to its core philosophy: as AI agents become more powerful and capable of destructive actions, ag-kit provides the safety rails — the "gravity" — that keeps agent actions controlled and bounded.

Features

• 20 domain specialist agent roles: Pre-defined agent personalities and expertise areas that can be activated by workflow context
• 47 progressive skills: A layered skill system with validation helpers that verify skill outputs before execution
• 13 repeatable workflow slash-commands: Standardized workflow templates for common agent tasks
• 6 workspace-wide rules: Global routing and safety rules that apply to all agent actions in the workspace
• Native PreToolUse safety gate: Blocks destructive commands (rm -rf /, drive formatting, raw disk overwrite patterns) while allowing safe operations (deleting dist/ or node_modules/)
• Merge-aware updates with rollback: Agent kits can be updated incrementally, and every update is reversible
• Plugin bundling with SHA-256 content inventory: Cryptographic verification of all bundled agent components
• MCP guidance: Built-in guidance for integrating Model Context Protocol servers

Installation

# Quick start with npx
npx @vudovn/ag-kit init

# Global install
npm install -g @vudovn/ag-kit
ag-kit init

Quick Start

# Initialize a new ag-kit workspace
npx @vudovn/ag-kit init

# The workspace is initialized with:
# - 20 specialist agent roles
# - 47 progressive skills
# - 13 workflow commands
# - 6 safety rules
# - PreToolUse safety gate (active by default)

Core Concepts

Antigravity-First Design

The "antigravity" philosophy inverts the conventional agent framework approach. Instead of starting with capabilities and adding safety later, ag-kit assumes safety is the foundation and capabilities are layered on top. This means destructive commands are blocked by default, and only explicitly permitted actions are allowed.

PreToolUse Safety Gate

The native PreToolUse gate is ag-kit's most distinctive feature. It intercepts every tool call the agent makes and evaluates it against a set of safety rules before execution. It blocks:

• Root filesystem deletion (rm -rf /, rm -rf \pwd``)
• Drive formatting commands
• Raw disk overwrite operations
• Any command matching known destructive patterns

And it allows:

• Deleting build artifacts (dist/, node_modules/)
• Normal file operations within the project
• Standard build and test commands

Progressive Skills

The 47 skills are organized progressively — foundational skills must validate before advanced skills become available. Each skill includes validation helpers that verify the skill's output meets quality and safety standards before it is considered complete.

Merge-Aware Updates

Agent kits can be updated incrementally. Each update is cryptographically signed with a SHA-256 content inventory, and every update is fully reversible with rollback support.

Advanced Features

Specialist Agent Roles

The 20 domain specialist roles include agents for code review, security analysis, documentation, testing, infrastructure, and more. Each role comes with predefined behavior patterns, tool access restrictions, and workflow hooks.

Workflow Slash-Commands

The 13 workflow commands provide standardized templates for common agent workflows:

• Code review workflows
• Security audit workflows
• Documentation generation
• Test suite generation
• Dependency analysis
• Migration planning

Plugin Bundling

Plugins are bundled with a SHA-256 content inventory that is verified on load, preventing supply chain attacks through compromised agent skills.

Pros

• ✅ Safety-first design with native PreToolUse gate
• ✅ Merge-aware updates with full rollback
• ✅ Cryptographic verification of all bundled components
• ✅ Broad skill library (47 progressive skills)
• ✅ Well-suited for teams that need auditability
• ✅ MIT license

Cons

• ❌ Newer and smaller ecosystem compared to established frameworks (CrewAI, LangGraph)
• ❌ Safety rules may be overly restrictive for some workflows
• ❌ Requires Node.js 22+ for the Antigravity tooling
• ❌ Smaller community and fewer integrations
• ❌ Documentation is still maturing

When to Use

ag-kit is ideal for:

• Teams building AI agent projects where safety and auditability are priorities
• Organizations that require rollback capability for agent kit updates
• Developers who want built-in safety guardrails without custom implementations
• Projects that need to comply with security policies around agent tool access
• Teams evaluating the trade-off between agent capability and agent safety

Resources

• GitHub Repository
• Documentation