NextLVLHasH

Lately I've been back deep in local AI agents. The app I'm building is designed to hook into a custom OS while still running on Windows — or honestly almost anything else. Paired with HasH-Arch, though, it turns into something else: a proper master's-level toolkit for structured cyber-security workflows, built on the BlackArch framework but heavily customised and improved for my own tooling.

For the tool documentation I leaned on Opus 4.7 with its 1M-token context and repo access to pull straight from the sources — then had it verify the details, map the Kali package data, and spot the naming differences between distros. On top of that I added my own explanations, corrections and workflow notes.

From there I built a custom agent workflow: Opus 4.7 carries out the heavier security-audit work, then hands tasks over to a locally-hosted 35B model that maps attack paths using GitHub access and the structured tool docs. The whole point is to see how useful a smaller model can be when it's given proper docs, tool maps and rules — instead of just being dropped into a terminal.

Still on the list: properly auditing the pentest agent and its tool-selection logic — whether it picks the right tools and stays inside a controlled workflow. But it's slowly starting to come together. Follow along on the roadmap and across the socials below.

It started with HelpmeGame — I built a game-playing AI just to understand, hands-on, how these models actually learn. Then I saw someone defining agents as plain files in Python, and it clicked: I built NextLVLAICoder, an agent that hooks into coding models and drives the computer the way Claude does — but on my terms.

Giving it real tools is where it got hard. Windows fought me at every turn; on Linux I tried Kali and BlackArch and hit constant instability, then landed on Arch — and went down the rabbit hole of building an entire distro just to make a single tool 100× more powerful. That distro is HasH-Arch.

And here's the thing: if I had to fight my own tools, most people will fight them harder — not everyone's put in the hours I have. So I built the stable, approachable version: the hacker's Windows — the full security toolkit, without the instability tax.

I've found 6 real bugs with AI-assisted hunting, and the big lesson was that no single model does it all. The answer is the right AI for the right job: Claude is brilliant at auditing and review but (rightly) won't write the offensive side — so NextLVLAICoder routes the audit to Claude and the heavier offensive-security generation to other models, all in an authorised, ethical-research context, with me owning the guardrails instead of the AI.