Built for testers who want to think, not babysit.
An autonomous pentest agent that thinks like an attacker. Point it at a URL, a repo, or a network and it runs the whole engagement itself—recon, exploitation, proof, remediation, and reporting—chaining 25+ specialized skills without a human stitching the steps together.
A live world model — every host, endpoint, and finding, linked as it scans.
What it does
25+ specialized skills spanning web, network, cloud, Active Directory, AI red-team, white-box source review, and threat modeling.
Detects a SQL injection, pivots to exploitation, writes the PoC—no manual stitching. The LLM decides what to run next based on what it just found.
50+ security tools run inside an ephemeral Kali Docker sandbox with hard cost, time, and call-count limits enforced server-side. The agent improvises; the blast radius doesn't.
Claude, OpenAI Codex, Gemini, Ollama, OpenRouter—or fully local on your own GPU via vLLM. The methodology ships with the skill; the model does the thinking.
Sandbox
Pick a target and watch Smith work — it fingerprints the surface, chains the right skills, and reports what it finds. Same brain, five very different targets.
One agent, five surfaces — each run picks and chains its own skills.
The world model
As Agent-Smith scans, it builds a knowledge graph of the target—hosts, ports, services, endpoints, technologies, credentials, and vulnerabilities, linked by how they relate. Findings reference the exact node they came from, coverage is tracked endpoint×technique, and the whole model persists to findings.json. That shared context is what lets skills chain intelligently instead of running blind.
How you run it
Give it a target—full engagement end-to-end, no human in the loop. Returns a report with every finding verified and a working PoC. Built for CI/CD, continuous testing, and dev self-serve.
A human expert drives strategy; Smith runs 50+ tools in parallel, tracks coverage, and writes the report—steer it live from the dashboard at localhost:7777. For high-value engagements and complex targets.
Quickstart
Requires Docker Desktop & Poetry. Run headless (above) or type claude to steer it yourself.
Open source, self-hosted, and model-agnostic. Clone it, point it at a target, and watch it work.