vxcontrol/pentagi

The problem

Penetration testing is slow, repetitive, and hard to staff. A tester has to scope a target, pick tools, chain their outputs, and write up findings, and much of that loop is mechanical. PentAGI puts a team of specialized agents in charge of that loop: you state a target and a goal in plain language, and the agents plan, run professional tools inside an isolated sandbox, keep notes, and produce a report. The pitch is not a new scanner but an orchestration layer over the tools testers already use.

What's different

Most automated pentest tooling is either a fixed scanner pipeline or a single-LLM wrapper around shell access. PentAGI's distinguishing choice is a multi-agent delegation model with a dedicated planner and execution monitoring, which the README argues matters when you run smaller or local models rather than top-tier cloud ones. Second, it is explicitly self-hosted and model-agnostic: a long provider list plus a custom OpenAI-compatible endpoint and a documented vLLM guide, so results never have to leave your infrastructure. Third, memory is persisted in PostgreSQL with pgvector and optionally a Neo4j knowledge graph, so successive flows can reuse earlier findings. Whether that agent orchestration beats a well-written manual playbook is exactly what the project has not demonstrated publicly.

Who it's for

Security engineers, pentesters, and red teams who already have authorization for their targets and want an assistant running on their own hardware, with their own model choice. It is a poor fit for anyone wanting a turnkey SaaS scanner, for compliance-driven engagements that require documented, human-signed testing steps, and for teams without the ops capacity to run a multi-service Docker Compose stack. Anyone looking for breach-and-attack simulation with predefined campaigns should look elsewhere; the README itself says that is not what this is today.

Cost of adoption

The install is a Docker Compose stack, and the README describes an interactive installer for Linux, macOS, and Windows. Expect multiple services: PostgreSQL with pgvector, the agent runtime, a web UI, and optional Neo4j, Grafana, and Prometheus, plus scraper and search-API keys if you want web intelligence. You also pay per token for whichever LLM you point it at, or pay in hardware and latency for local models. The license is MIT, but the README's most interesting pieces, such as agent profiles, plugins, MCP servers, and real-time attack-chain views, are framed as Enterprise add-ons. The README is long on setup and provider configuration but thin on evaluation data, so you will have to judge output quality yourself.

Verdict

PentAGI is roughly two years old and recently drew a sharp burst of attention: about 7,500 stars in one month on the Go board, against a quiet Hacker News thread with two points. That pattern, a large star count with little public technical discussion, argues for caution: the repo is real and actively documented, but the community has not yet stress-tested the claims. Treat it as a promising orchestration framework to pilot on an authorized lab target, not as a replacement for a tester's judgment. The Enterprise split is the main strategic risk if you build on the free tier.