🌌 AgentPM

❌ The Dark Age of Exchanging Prompts

As AI-assisted engineering (via Claude Code, Cursor, Windsurf) accelerates, developers are reduced to manually copy-pasting system prompts, rules, and "skills" from fragmented gists and unprotected repositories.

There is no standard architecture to manage, version, or secure custom AI instructions. Even worse, blindly copying third-party prompts introduces catastrophic Prompt Injection Vulnerabilities into your agentic workspace.

🌟 The Vanguard: AgentPM

AgentPM brings the robust npm and cargo experience directly to AI Agents. It transforms scattered text files into a scalable, secure, and standardized ecosystem.

Every skill installed via AgentPM is Cryptographically Verified via Sigstore, and passes through a Semantic Security Filter before executing. Any script (Bash/Python) is then run within an isolated E2B Cloud MicroVM Sandbox, completely protecting your local environment from prompt-injection execution payloads.

# 1. Initialize an AI-ready workspace
$ agentpm init

# 2. Securely install high-quality audited skills
$ agentpm install @oss/react-expert

🚀 Initializing install for skill: @oss/react-expert
🔍 Scanning deep abstraction layers for prompt injections...
✅ Skill audited: 0 malicious signatures found.
📦 Successfully deployed @oss/react-expert into .agents/skills/

✨ Features

| Feature | Description | |---------|-------------| | ⚡ Instant Discovery | Search 19,854+ production-grade AI skills from the AgentPM Registry | | 🛡️ Absolute Security | Containerized E2B MicroVM execution and Sigstore cryptographic signing | | 🏗️ Agentic Workspaces | agentpm init scaffolds .agents environments with structural integrity | | 🔍 Security Auditor | agentpm audit scans skill files for prompt injection and destructive commands | | 🤖 AI Skill Generator | agentpm generate creates new skills using offline templates or Gemini | | 📦 Publish & Share | agentpm publish contributes skills to the global registry | | 🖥️ MCP Server | agentpm mcp exposes tools via Model Context Protocol for AI agent integration | | 🌐 Platform Agnostic | Works with Claude Code (.agents), Cursor (.cursorrules), and Windsurf |

🚀 Quick Start

Installation

Install globally directly from the official NPM Registry:

npm install -g @amajumdar2249/agentpm

⚡ 1-Click AI Assistant Integration (MCP)

Auto-configure Claude Desktop, Cursor, and Windsurf to connect directly to AgentPM:

agentpm setup

Command Palette

# Scaffold an optimized agentic workspace
agentpm init

# Search the global registry (19,854+ skills)
agentpm search "react security"

# Securely install and audit a remote skill
agentpm install <skill-name>

# Audit a skill file for security threats
agentpm audit ./path/to/skill

# Generate a new skill from a prompt
agentpm generate my-skill "Build a REST API validator"

# View your orchestrated local environment
agentpm list

# Publish a skill to the registry
agentpm publish ./my-skill/

# Start the MCP server for AI agent integration
agentpm mcp

🏛️ Architecture

agentpm/
├── src/
│   ├── index.ts              # CLI entry point (Commander.js)
│   ├── scanner.ts            # 23+ regex patterns for prompt injection detection
│   ├── installer.ts          # Local-first, remote-fallback skill fetcher
│   ├── search.ts             # MiniSearch-powered fuzzy search engine
│   ├── schema.ts             # Zod-validated skill metadata schema
│   ├── mcp.ts                # Model Context Protocol server (4 tools)
│   └── commands/             # 10 CLI command handlers
├── registry/                 # Embedded registry (git submodule → agentpm-registry)
├── tests/                    # Jest test suite (8 test files)
└── web/                      # Next.js 16 registry browser (React 19)

🔗 Ecosystem

| Repository | Purpose | |------------|---------| | agentpm | CLI tool & MCP server (this repo) | | agentpm-registry | Global skill registry (19,854+ skills) | | devarmor | AI workstation security scanner (npm) |

🤝 Contributing

We welcome contributions to expand the AgentPM ecosystem. See CONTRIBUTING.md to get started.

🔒 Security

See SECURITY.md for our security policy and vulnerability reporting.

📄 License

MIT License — See LICENSE for details.