PentAGI

PentAGI is a self-hosted platform that runs autonomous AI agents to perform penetration testing tasks. It's built for security engineers, researchers, and teams who want an AI-driven pentesting workflow without giving up control over infrastructure or data. The system executes all actions inside sandboxed Docker containers, connects to a range of LLM providers, and stores results in PostgreSQL with pgvector for long-term memory.

Under the hood, PentAGI uses a multi-agent architecture: an orchestrator delegates work to specialized agents (researcher, developer, executor) that each handle a piece of the pentesting flow, from reconnaissance to exploitation to reporting. It ships with a GraphQL and REST API, a web UI, and integrations for observability (Grafana, Prometheus, Langfuse) and an optional Neo4j-backed knowledge graph (Graphiti) for tracking semantic relationships between findings.

Key features

• Sandboxed execution: every command and tool runs in an isolated Docker environment, separate from the host.
• Autonomous agents: a multi-agent system (orchestrator, researcher, developer, executor) plans and runs pentesting steps with optional monitoring and task planning.
• Built-in tool suite: 20+ professional security tools including nmap, metasploit, and sqlmap, pre-integrated for agent use.
• Memory and knowledge graph: long-term storage of research results in a pgvector-backed PostgreSQL store, with optional Graphiti/Neo4j integration for semantic relationship tracking.
• Web intelligence: a built-in scraper for browsing, plus integrations with Tavily, Firecrawl, Perplexity, Google Custom Search, Sploitus, Searxng, and others for information gathering.
• Flexible LLM support: works with OpenAI, Anthropic, Google AI/Gemini, AWS Bedrock, Ollama, DeepSeek, GLM, Kimi, Qwen, MiniMax, and custom/OpenAI-compatible endpoints, plus aggregators like OpenRouter and DeepInfra.
• APIs for automation: full REST and GraphQL APIs with Bearer token authentication for integrating PentAGI into other tooling.
• Monitoring stack: integrates with Grafana, Prometheus, Langfuse, VictoriaMetrics, Jaeger, and Loki for observability into both system and LLM behavior.
• Reporting: generates vulnerability reports with exploitation guides, viewable in-app, copyable, or downloadable as Markdown or PDF.
• Self-hosted and scalable: microservices architecture deployed via Docker Compose, supporting horizontal scaling and multi-instance setups.

Ideal use cases

PentAGI fits teams that want to automate parts of a penetration test, security researchers experimenting with agentic AI on offensive security tasks, and organizations that need a self-hosted alternative to cloud-only AI pentesting tools where data control matters. It's also useful if you want to compare how different LLM providers perform on the same pentesting workflow, since provider swapping is built in.

It's not a fit if you're looking for a Breach and Attack Simulation (BAS) product with predefined attack campaigns, like CALDERA-style adversary emulation; the README explicitly states this kind of scripted attack automation is conceptual, not implemented. It's also not suited for teams that need JSON export of flow reports, since that output format isn't currently supported (Markdown, PDF, and web view are). And because agents execute real tools against real targets, it requires proper authorization and isolation controls before pointing it at anything you don't own or have explicit permission to test.

Installation

PentAGI is designed to run via Docker Compose, with environment variables controlling LLM provider credentials, search API keys, and optional integrations (Langfuse, Graphiti, monitoring stack).

# clone the repository
git clone https://github.com/vxcontrol/pentagi.git
cd pentagi

# copy and edit the environment configuration
cp .env.example .env
# edit .env to add your LLM provider API key(s) (OpenAI, Anthropic, Ollama, etc.)
# and any optional search or monitoring integrations

# start the stack
docker compose up -d

After startup, the web UI is available for login, from which you can configure LLM providers, launch pentesting flows, and review reports. The project also documents multi-instance deployments (via TENANT_ID), Docker access controls for agents, and a guide for running local models through vLLM. Development and build instructions for contributors are included separately in the repository for those working on the Go backend or the React/TypeScript frontend.