| name | pentagi |
| description | PentAGI - Penetration testing Artificial General Intelligence. Fully autonomous AI pentest system. Use for: pentest, penetration testing, automated vulnerability scanning, security assessment. Docker-based with 20+ security tools, multi-agent architecture, knowledge graph memory. Located at ~/d/git/pentagi/. NOT for routine code review. |
PentAGI Skill
PentAGI — Penetration testing Artificial General Intelligence. Fully autonomous multi-agent pentesting system with sandboxed Docker execution, knowledge graph memory, and 20+ built-in security tools.
Location
~/d/git/pentagi/
Prerequisites
- Docker + Docker Compose must be running
- LLM API key (Anthropic, OpenAI, Ollama, etc.) set in
.env
- Minimum 16GB RAM recommended for full stack
Quick Start
cd ~/d/git/pentagi
cp .env.example .env
docker compose up -d
Architecture
Multi-agent system with specialist roles:
| Agent | Role |
|---|
| Orchestrator | Coordinates flow, queries memory, plans tasks |
| Researcher | Analyzes target, searches for vulnerabilities |
| Developer | Plans attack strategies, selects exploits |
| Executor | Runs security tools in sandboxed containers |
| Adviser/Mentor | Monitors execution, prevents loops (optional) |
| Reflector | Handles errors, graceful termination |
Key subsystems:
- Knowledge Graph — Neo4j + Graphiti for semantic memory
- Vector Store — PostgreSQL/pgvector for embeddings
- Security Tools — 20+ tools (nmap, metasploit, sqlmap, etc.) in sandboxed Docker
- Web Scraper — Isolated browser for live recon
- Monitoring — Grafana + Prometheus + Loki dashboards
Environment Variables
ANTHROPIC_API_KEY=
OPEN_AI_KEY=
EXECUTION_MONITOR_ENABLED=true
EXECUTION_MONITOR_SAME_TOOL_LIMIT=5
EXECUTION_MONITOR_TOTAL_TOOL_LIMIT=10
AGENT_PLANNING_STEP_ENABLED=true
API_BEARER_TOKEN=your-token-here
Key Features
- Execution Monitoring — Beta feature. Adviser agent intervenes when it detects loops or stuck execution. 2x quality improvement on smaller models (<32B). Enable with
EXECUTION_MONITOR_ENABLED=true
- Task Planning — Beta feature. Planner decomposes complex attacks into 3-7 actionable steps before execution. Enable with
AGENT_PLANNING_STEP_ENABLED=true
- Smart Memory — Learns from past pentests. Stores successful approaches for reuse
- Chain Summarization — Manages context window growth automatically
Comparing to Other Tools
| Tool | Scope | Model | Best For |
|---|
| pentagi | Full pentest, multi-agent, 20+ tools | Any LLM | Production autonomous pentesting |
| shannon | Full pentest, 5-phase | Claude only | White-box analysis + POC exploits |
| argus | Reconnaissance | N/A | Quick surface scans |
| webcopilot | Lightweight scan | N/A | Known CVE patterns |
PentAGI is the heaviest option — full Docker stack, Neo4j, PostgreSQL, multi-agent orchestration. Use when you need fully autonomous pentesting with memory and planning.
Tips
- For small models (<32B): enable both Execution Monitoring and Task Planning for 2x better results
- Adviser agent works best with a stronger model or max reasoning mode
- First run:
docker compose up -d pulls ~5-10GB of images. Wait before accessing UI
- API docs: Swagger at
http://localhost:8080/swagger/ after startup
- Reports generated per pentest flow — check the UI or API for artifacts