# PentAGI with an abliterated LLM

> Pointing PentAGI's custom LLM provider at Cyber v1, Nodecode's abliterated model: four lines of .env, and the model field that makes them work.

Source: https://nodecode.ai/docs/cyber-pentagi
Updated: 2026-11-10

[PentAGI](https://pentagi.com) is a self-hosted multi-agent pentest system: a team of agents plans the work, runs nmap, Metasploit, sqlmap and 20 more tools inside a Docker sandbox, and keeps what it learned in PostgreSQL. It takes any OpenAI-compatible endpoint through one block of `.env`.

## Install it

Its installer sets up the whole stack with Docker Compose: download the archive for your architecture from [pentagi.com](https://pentagi.com), unzip it, and run the installer (it needs the Docker socket, so `sudo ./installer`).

```sh
wget -O installer.zip https://pentagi.com/downloads/linux/amd64/installer-latest.zip
unzip installer.zip
sudo ./installer
```

## Point it at Cyber v1

The installer writes a `.env`. Its custom provider is the block that takes an arbitrary base URL, a key and a model name:

```sh
LLM_SERVER_URL=https://api.nodecode.ai/v1
LLM_SERVER_KEY=sk-nc-...
LLM_SERVER_MODEL=nodecode-cyber-v1
LLM_SERVER_API_TYPE=
```

Four variables, and the empty fourth matters: `LLM_SERVER_API_TYPE` is `azure` or `azure_ad` for an Azure deployment, and empty for a plain OpenAI-compatible endpoint, which is what this is. Leave `LLM_SERVER_PROVIDER` empty too: it prefixes model names for a LiteLLM proxy (`moonshot/kimi-2.5`), and `LLM_SERVER_MODEL` is already the name the API answers to.

A reasoning model keeps its thinking across turns when the request asks it to. If PentAGI's chains come back with the reasoning stripped, set `LLM_SERVER_PRESERVE_REASONING=true`, which is the switch its README names for providers that require it.

Then bring the stack up again so the new `.env` is read:

```sh
docker compose up -d --force-recreate
```

## The other route

PentAGI's OpenAI provider takes a base URL of its own, so `OPEN_AI_KEY=sk-nc-...` with `OPEN_AI_SERVER_URL=https://api.nodecode.ai/v1` also lands on Cyber v1. The custom provider above is the one to prefer: it names the model directly and leaves the OpenAI defaults alone.

## Choosing the model per agent

`LLM_SERVER_CONFIG_PATH` takes a YAML file that gives each of PentAGI's agents its own model. One model for every agent is the default, and the model to start with; a second, smaller model for the parser and the summariser is the usual first cut.

## What was checked

The settings above are PentAGI's own, as its README and `.env.example` give them, and the endpoint they point at answers as [the quickstart](https://nodecode.ai/docs/cyber) shows. PentAGI's whole stack has not been run against it here yet.