What it does
PentAGI is an open-source, self-hosted system in which an AI agent works through penetration testing tasks. You define the task, the agent analyses it and decides on an approach, then carries out the testing itself using a terminal, a browser, an editor and external search, working out each next step as it goes. Everything runs in a sandboxed Docker environment that includes more than 200 Kali Linux penetration testing tools. Commands and their output are saved in PostgreSQL, a knowledge graph built on Graphiti and Neo4j keeps semantic memory of the work, and Langfuse integration lets you monitor the agents. A web UI shows the testing as it unfolds and presents results and reporting at the end. The site lists twelve or more LLM providers, including OpenAI, Anthropic, Google Gemini, AWS Bedrock, DeepSeek and local models through Ollama, plus any OpenAI-compatible endpoint. A REST API and GraphQL let you drive the system from outside, and an observability stack with tools such as Grafana, Loki, Jaeger and OpenTelemetry is supported.
Capabilities
- Sandboxed Docker environment with 200+ Kali Linux penetration testing tools
- Works out its next step itself and can search the web for current information
- Knowledge graph memory (Graphiti and Neo4j) and command history in PostgreSQL
- Web UI to watch testing in real time, plus REST and GraphQL APIs
- Twelve or more LLM providers, including local models through Ollama
- Langfuse integration and an observability stack for monitoring the agents
When to use it
- A security team wants an AI agent to take on penetration testing tasks on infrastructure it runs itself
- You want every command and output from a test saved in PostgreSQL as a history of the work
- You prefer to keep testing private by pairing it with local models through Ollama
