🦄 Pownie

I vibe-hacked my way to Top #87 Global on Hack The Box. Hall of Fame. Built entirely on Claude Code.

Pownie is the harness that got me there, a Claude Code plugin that wires up persistent intel, lifecycle hooks, and multi-agent coordination for offensive security.

The model already knows how to hack. It doesn't need playbooks, it needs hands and legs that lets its knowledge compound over long engagements. That's what this is.

What it does:

• Hooks fire on every tool call — auto-extract credentials, log attempts, surface prior intel before the model repeats itself
• Neo4j intel graph stores everything the model discovers, outside the context window, where compaction can't reach it
• Attack class tracking kills entire categories of attack when evidence shows they're impossible on the target
• Multi-agent orchestration spawns parallel teammates after recon or shell access
• Context survival — PreCompact hook snapshots state to Neo4j before compaction wipes the window

It battle-tested across hundreds of HTB machines over 2 months. From #9000+ to Hall of Fame.

---

⚠️ Required MCP Servers

This plugin does not work standalone. It requires a specific MCP server stack to function. Don't worry, just run ./setup.sh and it builds and starts everything for you.

MCP Server	Purpose	Required
mcp-kali	Kali Linux command execution, sessions, background tasks, proxy	Yes
neo4j-mcp	Knowledge graph for attack state, credentials, attempt tracking	Yes
playwright	Browser automation with headed Chromium + noVNC	Optional

---

✨ What the Plugin Adds

Component	What It Does
Intel Graph	Neo4j-backed knowledge graph — tracks targets, credentials, services, vulnerabilities, shells, and flags as structured data that survives context compaction
Pre/Post Hooks	Automatic tracing on every Kali MCP call — logs commands to Neo4j, extracts credentials and services from output, surfaces prior attempts before execution
Strategic Compaction	Context management for long offsec sessions — phase-aware compaction with Neo4j state preservation and rich recovery files
Multi-Agent Orchestration	Spawns parallel teammates after recon or shell access — CVE scouts, code analysts, system enumerators working concurrently
Debrief & Writeup	Post-challenge writeup generation with structured failure analysis and MEMORY.md updates

---

🚀 Quick Start

Prerequisites

• Docker (with Docker Compose)
• Claude Code v1.0.33+

1. Clone and setup the stack

git clone https://github.com/d0gesec/pownie.git
cd pownie
./setup.sh

The setup wizard will:

• Ask which optional services to enable (browser, telemetry)
• Detect container name conflicts and offer alternatives
• Generate docker-compose.yml, .mcp.json, and .claude/settings.local.json
• Build Docker images and start the stack
• Wait for Neo4j to be healthy

2. Install the plugin

Option A — Marketplace install:

/plugin marketplace add d0gesec/pownie
/plugin install pownie@d0gesec

Option B — Direct from cloned repo:

claude --plugin-dir ./pownie

3. Start hacking

The plugin activates automatically. Skills like the intel graph and strategic compaction work in the background. User-invocable skills:

• /pownie:offsec-debrief — generate writeup after completing a challenge

---

🎯 Basic Workflow

1. Give it a target

Tell Claude the ctf target and the goal. That's it.

CTF target's IP 10.10.11.42. Capture the user flag.

The plugin handles the rest in the background — hooks fire on every tool call, credentials get extracted automatically, attempts get logged, and the intel graph builds itself as the model works.

2. Spawn teammates for layered attacks

After the initial enumeration, kick off the orchestrator to throw multiple approaches at the target in parallel.

/pownie:offsec-lead

This triggers the offsec-lead skill, which spawns 2-3 parallel Agent teammates for CVE scouts, code analysts, system enumerators, each bootstrapping from the same Neo4j intel graph. Same credentials, same disproven attack classes, no duplicate work.

Note: Multi-agent coordination is currently experimental and disabled by default. I usually engage it after the enum stage when there are multiple attack surfaces to explore concurrently.

3. Debrief

After capturing flags, generate a structured writeup with failure analysis.

/pownie:offsec-debrief

---

🔧 Setup Options

./setup.sh              # interactive — choose components
./setup.sh --all        # everything (browser + telemetry)
./setup.sh --core-only  # just kali + neo4j
./setup.sh --bare       # bare Kali/Linux mode (no mcp-kali container)
./setup.sh --down       # stop containers

./cleanup.sh            # stop containers, keep data
./cleanup.sh --volumes  # also wipe neo4j data and workspace
./cleanup.sh --full     # remove everything including generated files

Bare mode

If you're running Claude Code directly on a Kali/Linux machine instead of through mcp-kali containers:

./setup.sh --bare

This skips the kali container, sets hook matchers to fire on Bash tool calls, and only spins up Neo4j in Docker. You still get the full intel graph, strategic compaction, and all skills.

---

📐 Architecture

┌─────────────────────────────────────────────────────┐
│  Claude Code + pownie plugin                        │
│                                                     │
│  skills/          hooks/           .mcp.json        │
│  ├─ intel-graph   ├─ PreToolUse    ├─ kali (stdio)  │
│  ├─ compact       ├─ PostToolUse   ├─ neo4j (stdio) │
│  ├─ offsec-lead   └─ PreCompact    └─ playwright    │
│  ├─ debrief                            (http)       │
│  └─ debrief                                         │
└──────────┬──────────────┬──────────────┬────────────┘
           │              │              │
     ┌─────▼─────┐  ┌────▼────┐  ┌──────▼──────┐
     │ pownie-   │  │ pownie- │  │  pownie-     │
     │ kali      │  │ neo4j   │  │  playwright  │
     │           │  │         │  │              │
     │ mcp-kali  │  │ neo4j   │  │ @playwright/ │
     │ server    │  │ + mcp   │  │ mcp + noVNC  │
     │ 1000+     │  │ bolt    │  │              │
     │ sec tools │  │ :7687   │  │ :3888 :6080  │
     └───────────┘  └─────────┘  └──────────────┘

How the hooks work

Every Kali MCP tool call flows through the hook pipeline:

1. PreToolUse (pre-exec.sh) — queries Neo4j for prior attempts, known credentials, disproven attack classes, and active strategies on the target IP. Surfaces this as context so the agent avoids repeating failed approaches.

2. Agent executes command via mcp-kali

3. PostToolUse (post-exec.sh) — logs the command and result to Neo4j, auto-extracts credentials and services from output, prompts for phase classification, detects shell acquisition and repeated failures.

4. PreCompact (pre-compact-save.sh) — before any context compaction, queries Neo4j and writes a rich compact-state.md with targets, credentials, failed attempts, and command history for post-compaction recovery.

---

🗃️ Plugin Structure

pownie/
├── .claude-plugin/
│   └── plugin.json              # plugin manifest
├── skills/
│   ├── offsec-intel-graph/      # neo4j knowledge graph schema + usage
│   │   ├── preToolUse/pre-exec.sh  # pre-execution context retrieval
│   │   └── postToolUse/post-exec.sh # post-execution logging + intel extraction
│   ├── strategy-compact/        # context compaction strategy
│   │   └── pre-compact-save.sh  # neo4j state snapshot before compaction
│   ├── offsec-lead/             # multi-agent orchestrator
│   └── offsec-debrief/          # writeup generation
├── hooks/
│   └── hooks.json               # event hook wiring
├── docker/                      # build-from-source Dockerfiles
│   ├── Dockerfile.kali
│   ├── Dockerfile.neo4j
│   ├── Dockerfile.playwright
│   └── ...
├── setup.sh                     # interactive setup wizard
├── cleanup.sh                   # teardown script
├── LICENSE
└── README.md

Generated at runtime by setup.sh (gitignored):

• docker-compose.yml
• .mcp.json
• .claude/settings.local.json

---

🔍 Telemetry (Optional)

Enable the telemetry profile during setup to get Grafana + Tempo for trace visualization. Useful for debugging skills and hooks during development.

./setup.sh  # answer 'y' to "Enable trace viewer"

Then open Grafana at http://localhost:3000 → Explore → Tempo → search by service.name = pownie-pre-hook.

The hooks emit OTLP spans for every Kali MCP call regardless — telemetry just gives you a UI to browse them. Without it, spans are silently dropped with zero impact on functionality.

---

⚠️ Disclaimer

This project is shared for educational and authorized security testing purposes only. It orchestrates unrestricted command execution inside a Kali Linux container — use it responsibly and at your own risk. The authors assume no liability for misuse. Always ensure you have proper authorization before testing any target.

---

📄 License

MIT