Agentic security tooling for your own team.

Agenti is a free, self-hosted tool that lets a security team run AI agents for network enumeration, OSINT, traffic analysis and offline hash cracking. It runs on your own Kali Linux machine, built on the Claude Agent SDK. An administrator approves the scope, and every action is checked against it and logged.

Download Read the docs Security

What it is

Agenti is built for one organisation's internal security team. You install it on a Kali Linux machine. A team member picks an agent, gives it a target and a model, and follows its work from a web page or the command line. The agent plans its own steps, runs reviewed programs, reads the results and writes a report.

It is not a hosted service. You run it and you keep the data. There are two roles: operator, who runs agents, and admin, who sets the scope, approves it, and manages users and agents.

How a run works

An admin records the engagement: the client, the approved targets, the time window, the allowed models and the spending limits, then approves it.
An operator starts an agent with a target and a model. A target can only narrow the approved scope.
The agent works through a fixed list of reviewed programs. Each command is checked against the scope before it runs.
The operator can steer. Messages redirect the agent's effort; they do not give it new permissions.
A report is written with every command, every operator message and every attached file.

The agents

enum

Network enumeration against an approved address range.

research

Open-source intelligence against an approved domain.

trafficmon

Passive packet capture, analysed and summarised.

cracker

Offline hash cracking on files an operator attaches. No network access.

Build your own

Admins can build recon and file-analysis agents in settings. A built agent is configuration, not new code.

exploit off by default

An optional agent that stays disabled until an administrator enables it, with extra approval steps.

Built to be checked

We describe Agenti as auditable, bounded, isolated, and admin-authorised for real targets. We do not call it "safe": a tool like this is only as safe as the people configuring it.

The scope is enforced by the system, not by the AI model.
Everything runs without administrator privileges.
Data is encrypted at rest.
Decisions, runs and changes are logged.

Read the security page, including what Agenti does not protect against.

Requirements

Kali Linux, or another Debian-based system, with systemd.
An Anthropic API key.
Root access to install and to unlock the encrypted vault.

Licence: to be announced with the first release. Source: github.com/agenti-foss/agenti.