PentAGI

Fully autonomous, self-hosted penetration testing system powered by multi-agent AI. OSS (GitHub: vxcontrol/pentagi). Docker Compose deployment. Supports 10+…

PentAGI

Fully autonomous, self-hosted penetration testing system powered by multi-agent AI. OSS (GitHub: vxcontrol/pentagi). Docker Compose deployment. Supports 10+ LLM providers including Anthropic.

What It Does

PentAGI runs autonomous pen tests end-to-end: recon, exploitation planning, tool execution, memory storage, and report generation. Human sets the target; agents drive the rest with optional monitoring.

Agent Architecture

Three-tier delegation:

Agent	Role
Orchestrator	Plans flow, routes to specialists, synthesizes results
Researcher	Gathers intel, queries knowledge base, identifies vulnerabilities
Developer	Plans attack sequences, selects exploits
Executor	Runs tools, stores results, reports back

Data model: Flow → Task → SubTask → Action → Artifact/Memory. Each action produces artifacts (files, reports, logs) and memories (observations stored as vectors).

Memory System

Three memory tiers:

• Long-term: vector store (pgvector), knowledge base, tool usage patterns
• Working: current task context, active goals, system state
• Episodic: past actions, outcomes, success patterns

Orchestrator queries vector store at flow start to surface similar past tasks — avoids re-exploring known terrain.

Context Management

Built-in chain summarization: when conversation chain grows too large, converts to ChainAST, summarizes older message pairs, rebuilds a smaller chain. Configurable: preserve-last-section (default 50KB), max body pair size (16KB). This prevents token limit failures mid-pentest.

See Context Compression Strategies — PentAGI’s chain summarization is an instance of anchored iterative summarization.

Security Tools (20+)

Includes: nmap, sqlmap, metasploit, and web browser (isolated scraper container). All tools run in sandboxed Docker containers — host is never directly exposed. Tool selection is automatic based on task requirements.

Infrastructure Stack

• Backend: Go + GraphQL API
• Frontend: React + TypeScript
• Storage: PostgreSQL + pgvector (vector search), Neo4j + Graphiti (knowledge graph)
• Observability: Grafana, VictoriaMetrics, Jaeger (tracing), Loki (logs), Langfuse (LLM analytics)
• Queue: async task queue between API and agent layer

Key Differentiators vs. Our Design

Feature	PentAGI	Our pentest-agent
Scope	General-purpose, any target	Next.js + ECS + Neon specific
Memory	pgvector + Neo4j knowledge graph	Shared filesystem (JSON)
Tools	20+ (metasploit, etc.)	Targeted: nuclei, nmap, sqlmap, testssl, Prowler, Trivy
Deployment	Docker Compose microservices	Local script → GitHub Actions
Context mgmt	Chain summarization	Per-agent isolated context
Auth handling	Not highlighted	Dedicated test account + session injection

What to Borrow

• Flow → Task → SubTask hierarchy maps to our supervisor → specialist → tool call structure
• Accessible vs. target domains — same as AWS Security Agent’s pattern
• Chain summarization — relevant if our agents hit context limits during long scans
• Sandboxed tool execution — run CLI tools in subprocess isolation, not direct shell

Related Pages

• Aws Security Agent — managed AWS alternative; compare scope control patterns
• Pen Test Agent Design — our custom agent built from both reference sources
• Agent Harness — supervisor pattern PentAGI instantiates
• Context Compression Strategies — PentAGI’s chain summarization is a live example