PentAGI: hack-test your own app + sell security audits to clients

PentAGI is made by VXControl, a security engineering company in Dubai. It is MIT licensed, runs on your own machine, and had 33,157 stars on GitHub on 3 November 2026. You tell it what to test in plain words. Its agents do the hacking.

Only test apps you own or have written permission to test.

That line comes first because the tool is real. It runs the same tools attackers use. PentAGI's own EULA allows it only for "authorized penetration testing" where you have explicit permission from the owner. More on this in section 02.

Read this guide: https://codewithnishant.dev/free/pentagi-ai-pentesting/ and the official PentAGI README at https://github.com/vxcontrol/pentagi. Help me set up PentAGI on this machine to test my own app. First check that Docker and Docker Compose are installed and that I have at least 2 vCPU, 4 GB RAM and 20 GB free disk. Then walk me through the official installer for my operating system, downloaded only from the pentagi.com links in the README. Ask me which AI provider to use: Anthropic Claude with my own API key, or a local model through Ollama. Never print my API keys. After it starts, remind me to change the default admin password. Only help me test apps I own or have written permission to test, and point it at a staging copy, not my live app.

01. What PentAGI is

PentAGI stands for Penetration testing Artificial General Intelligence. The README calls it "an open-source, self-hosted multi-agent system for autonomous penetration testing".

PentAGI is not an Anthropic product. It can use Claude as its brain, the same way it can use other models.

How a run works:

• You describe it. A target, a goal and the limits, in plain English.
• The agents plan and attack. They pick tools, run them in the sandbox and search the web for known exploits.
• It remembers. Results and approaches that worked are stored for later runs.
• You get a report. The README promises "thorough vulnerability reports with exploitation guides".

02. The one rule: own it or get it in writing

The README says it plainly: "Only test systems you own or are explicitly authorized to assess."

Also check the hosting provider's own testing rules before you test an app on someone's servers.

03. What you need

From the README's system requirements:

• Docker and Docker Compose (Podman also works)
• At least 2 vCPU
• At least 4 GB RAM
• 20 GB free disk
• Internet, to pull the images and tools
• One AI provider: a Claude API key, or a local model through Ollama

For real client work, the README recommends a two-node setup that runs the attack containers on a separate server. For your first test, one machine is fine.

04. Install PentAGI

Two ways, pick one

Way A: the installer (recommended)

PentAGI ships an interactive installer for Linux, Windows and macOS. Download the zip for your machine from the links in the README's Quick Start:

• Linux: amd64 or arm64
• Windows: amd64
• macOS: Intel (amd64) or M-series (arm64)

The README's commands for Linux amd64:

mkdir -p pentagi && cd pentagi
wget -O installer.zip https://pentagi.com/downloads/linux/amd64/installer-latest.zip
unzip installer.zip
./installer

On a Mac or Windows, it is the same idea: download your zip, unzip it, run the installer. If macOS flags the file, use only the official links and follow the README's installer troubleshooting note.

Docker permission. The installer talks to Docker, so either run it with sudo ./installer or add your user to the docker group. The README warns the docker group is "root-equivalent", so only do that on a machine you trust.

What the installer does for you:

1. Checks Docker, the network and your system
2. Creates the .env settings file
3. Sets up your AI provider (Claude, Ollama and the rest)
4. Sets up search engines (DuckDuckGo and others, optional)
5. Generates secure passwords and SSL certificates
6. Starts PentAGI with Docker Compose

Way B: Docker Compose by hand

The README's manual steps, in short:

mkdir pentagi && cd pentagi
curl -o .env https://raw.githubusercontent.com/vxcontrol/pentagi/master/.env.example
curl -o example.custom.provider.yml https://raw.githubusercontent.com/vxcontrol/pentagi/master/examples/configs/custom-openai.provider.yml
curl -o example.ollama.provider.yml https://raw.githubusercontent.com/vxcontrol/pentagi/master/examples/configs/ollama-llama318b.provider.yml
curl -o example.bedrock.provider.yml https://raw.githubusercontent.com/vxcontrol/pentagi/master/examples/configs/bedrock.provider.yml

Then open .env and:

• Add at least one AI provider (next section).
• Change the security values, like COOKIE_SIGNING_SALT and the Postgres user and password.

Start it:

curl -O https://raw.githubusercontent.com/vxcontrol/pentagi/master/docker-compose.yml
docker compose up -d

First login

• Open https://localhost:8443 in your browser.
• A fresh install has one admin account: [email protected] with password admin.
• Change that password first. The README says to do it before any real work.

05. Pick the brain: Claude or a local model

PentAGI needs at least one AI provider. Two good starting points:

Run it on Claude

Get an API key from the Claude Console and put it in .env (or paste it into the installer):

ANTHROPIC_API_KEY=your_anthropic_api_key

The README's Anthropic section lists the prices it uses, per 1M tokens: claude-sonnet-5 at $3 input and $15 output, claude-haiku-4-5 at $1 input and $5 output. Check Anthropic's own pricing page before you quote a client, prices change.

Run it on a free local model

The README's basic local setup, from its Ollama section:

OLLAMA_SERVER_URL=http://localhost:11434
OLLAMA_SERVER_MODEL=llama3.1:8b-instruct-q8_0

The README says Ollama "provides zero-cost local inference if you have sufficient computational resources". Two tips from it:

• Small models need help. For models under 32B parameters, turn on EXECUTION_MONITOR_ENABLED=true and AGENT_PLANNING_STEP_ENABLED=true. The README reports about 2x better results, with 2 to 3x more time and tokens.
• Bigger local setups exist. The README links a full guide for vLLM with Qwen3.5-27B for fully local work.

You pick the provider again for each test, in the New Flow form. Settings, then Providers, lets you choose models per agent.

06. Your first test on your own app

Start small, one target, one goal

1. Make a copy. Run your app on staging or locally, with test data only.
2. Open Flows, then New Flow.
3. Pick a mode. Automation runs the whole test end to end. Assistant lets you steer step by step, and its Use Agents toggle hands parts to specialist agents.
4. Pick your provider. Claude or your local model.
5. Write the goal. Target, type of test, limits and the result you want. Example below.
6. Watch it work. The flow page shows the agents' messages, tool activity, terminal output and tasks. The Assistant view can stop or redirect the run.
7. Get the report. The Report menu opens it in the browser, copies it, or downloads it as Markdown or PDF.

A first goal, based on the README's own example:

Assess https://staging.yourapp.com for common web application vulnerabilities. Focus on authentication, file handling and injection issues. Stay within the provided target only and summarize confirmed findings with reproduction steps.

Want a deeper test? The repo has a base web pentest prompt that walks through every endpoint and checks path traversal, CSRF, XSS, SQL injection with sqlmap and command injection. Save the one you use as a Template, so the next run takes one click.

After the report: fix the issues, then run the same template again to confirm they are gone. Pair it with my vibe coding safety checklist and the 30 security holes your AI never checks for the things a scan cannot see.

07. Sell it: security audits as a service

If you build apps for clients, every app you ship is a security question waiting to be asked. PentAGI lets you answer it with a report. Here is a simple way to package that.

Who buys it: clients you already built apps for, small businesses with an online store or booking app, and founders about to launch.

Step 1: Agree the scope

• The exact URLs and environments (staging or live).
• Test accounts they create for you.
• What is out: payment gateways, email providers, anything run by a third party.
• The dates and hours you will test.

Step 2: Get written permission

A signed letter or email from the person who owns the app. It names you, the scope, the dates, your contact, and what you do if you find personal data. No signature, no scan. For a big client, have a lawyer check the wording. This is a practical checklist, not legal advice.

Step 3: Run the test and check every finding

Run PentAGI inside the scope. Then check each finding by hand before it reaches the client. An AI agent can be wrong, and one false alarm in a report costs you trust.

Step 4: Hand over a report with fixes

• A short summary in plain words for the owner.
• Each finding with severity (critical, high, medium, low), proof, and how to fix it.
• An offer to fix it. If you built the app, you are the right person to patch it, as paid work.

The PentAGI report is your raw material, not the final deliverable. Rewrite it for the client (prompt in section 09).

Step 5: Retest every month

Apps change every week. New features bring new holes. Offer a monthly retest of the same scope, with a short "what changed" report. Your saved template makes the rerun quick.

Pricing: set your own. A clear structure works: one price for the first audit, then a monthly retainer for retests and fixes.

Be honest about what it is. Sell it as a regular security check. A client who needs a pentest for a compliance certificate may need a certified firm. And the EULA lets you use PentAGI for authorized testing, not resell the software itself. You are selling your work.

New to finding clients? Start with my first client playbook, and read why selling one clear service beats selling everything.

08. Where it goes wrong

• Testing live production. Use a staging copy unless the client agrees in writing.
• Leaving admin / admin. Change the default password at first login.
• Open sandbox ports. Each test publishes two ports from 28000 to 29999 on every network interface by default. On a server, block that range with a firewall, the README explains how.
• Trusting the report blindly. Verify every finding yourself.
• Scope creep. Write "stay within the provided target only" into every goal.
• A surprise model bill. Long runs use a lot of tokens, more with monitoring and planning on. Watch your provider's usage page.
• Client data in a cloud model. The EULA puts data protection on you. For sensitive clients, use a local model.

09. Copyable prompts

1. Scope and permission email to a client

Hi [name], before I run the security check on [app name], I need your written OK. Scope: [URLs], environment: [staging or live], test accounts: [list], out of scope: [payment gateway, email provider, other third parties], dates: [from to], my contact: [phone]. If I find personal data, I stop, tell you, and do not copy it. Please reply "Approved" from your company email, or sign the attached letter.

2. Turn the PentAGI report into a client report (paste into Claude with the Markdown report)

Here is a PentAGI report in Markdown. Rewrite it for a non-technical business owner. Start with a 5-line summary. Then list each finding with: severity (critical, high, medium, low), what it means for their business in one line, the proof, and the fix. Mark any finding I have not confirmed by hand as "to verify". Do not add findings that are not in the report.

3. Monthly retest goal

Retest https://staging.yourapp.com within the same scope as last month. First recheck these findings from the last report: [list]. Then test the new features: [list]. Stay within the provided target only. Report which old issues are fixed, which remain, and any new confirmed findings with reproduction steps.

FAQ

What is PentAGI?

PentAGI is a free, open source, self-hosted AI tool for penetration testing, made by the security company VXControl. You describe a target and a goal in plain words, and a team of AI agents runs 20+ security tools such as nmap, Metasploit and sqlmap inside an isolated Docker sandbox, then writes you a report.

Is PentAGI free?

Yes. The code is MIT licensed and you host it yourself. The only running cost is the AI model: a cloud provider like Anthropic bills you per token, while a local model through Ollama costs nothing beyond your own hardware.

Can PentAGI run on Claude?

Yes. Anthropic is one of its built-in providers. Put your key in ANTHROPIC_API_KEY in the .env file, or pick Anthropic in the installer. By default PentAGI uses claude-sonnet-5 for the main agents and claude-haiku-4-5 for the lighter ones.

Is PentAGI made by Anthropic?

No. PentAGI is made by VXControl, a security engineering company based in Dubai. It can use Anthropic's Claude models as its brain, along with OpenAI, Gemini, AWS Bedrock, local models and others.

Can PentAGI run with a free local model?

Yes, through Ollama or vLLM. Set OLLAMA_SERVER_URL and OLLAMA_SERVER_MODEL in the .env file. You need a machine strong enough for the model, and for models under 32B parameters the README says to turn on execution monitoring and task planning.

What do I need to run PentAGI?

Docker and Docker Compose (or Podman), at least 2 vCPU, 4 GB of RAM, 20 GB of free disk and internet access, plus one AI provider key or a local model.

Is it legal to use PentAGI?

Only on systems you own or have explicit permission to test. Its EULA allows lawful, authorized penetration testing only. Scanning or attacking someone else's app without permission is a crime in most countries.

Can I sell security audits using PentAGI?

You can sell your own service: scoping, testing with written permission, checking every finding by hand, a report with fixes and regular retests. The EULA does not let you resell the PentAGI software itself, and a client who needs a certified pentest for compliance may need a certified firm.

What I checked, and what I did not

On 3 November 2026 I read the official README, the EULA and the example web pentest prompt, and checked the repo on the GitHub API: 33,157 stars, 5,179 forks, MIT, last push 22 October 2026. Every command above is copied from the README. I did not install or run PentAGI myself, so the setup steps are the maker's, not a test of mine. Thanks to VXControl for building it in the open.

Hack your own app before someone else does. Then charge clients to do the same for theirs.