Hexbane Try it live
Web · Python · GitHub Actionlive

Check an AI agent skill or MCP server before you trust it

Finds hidden instructions, tool poisoning, secret theft, download-and-run commands and risky MCP settings - in your browser, the CLI or a GitHub Action.

Web · Python · GitHub Action
hexbane · live demolive demo
Hexbane screenshot
3platforms: Web, Python, GitHub Action
11/11held-out attacks flagged
0/5held-out false alarms
65/65disguised attacks flagged (53 before unfolding)
Why it exists

Break the hex before your agent falls for it.

AI agents now load skills, MCP servers and tools written by strangers.

Hexbane checks those files before your agent trusts them, shows what any hidden or disguised text really says, then guards what the agent does while it runs.

What's inside

What the gates check.

Load gate

Paste or drop a SKILL.md, AGENTS.md, .cursorrules, mcp.json, tool list or a whole skill folder. Runs in your browser; files are never uploaded.

Hidden and disguised text

Finds zero-width and tag characters, bidi overrides, HTML-comment and base64 payloads, and unfolds look-alike Cyrillic/Greek letters, entities, s p a c e d words and split words - then shows the decoded text.

Tool poisoning

Flags 'ignore previous instructions', 'do not tell the user', fake <IMPORTANT> tags, shadowing of other tools and requests to pass secrets as parameters.

Risky MCP config

Unpinned npx/uvx packages, shell wrappers, plain-HTTP remotes, privileged Docker, inline API keys and broad filesystem roots.

Rug-pull lock

`hexbane lock` pins every file and tool description; a later scan with --lock flags anything that changed after you approved it.

Call gate

allow / ask / deny for each tool call (rm -rf /, reads of ~/.ssh, pipes to a shell) and quarantine for tool results that carry injected instructions.

Chat gate

From ScopeGuard: keeps a business assistant on its mission - allow, steer back, redirect, throttle off-topic use, block jailbreaks.

Fixes you can paste

Every finding has a copy-paste fix and a ready 'reviewed' suppression line; `hexbane fix` strips invisible characters. A GitHub Action puts findings in code scanning (SARIF).

Run it yourself

Scan a folder in one line.

From the README’s “Quick start” section. Full options, requirements and troubleshooting are there too.

Open the README
bash
$ git clone https://github.com/juw12/hexbane && cd hexbane
$ python -m hexbane scan path/to/skill-or-repo
$ python -m hexbane scan . --format sarif --out hexbane.sarif
$ python -m hexbane lock path/to/skills
Your data

Nothing leaves your machine

  • T
  • h
  • e
Good to know

Honest limits.

Straight from the README, so you know what you're getting before you try it.

  • A
  • d
  • e
  • t
Under the hood

How it's built.

Built with

Language
PythonJavaScriptHTML/CSS
Platform
PWAOffline service worker
Testing
Automated tests

69 files in the repo · last updated 10 Nov 2026.

Browse the code
Questions

Good questions.

What does it run on?

Web, Python, GitHub Action. You can try it in the browser with nothing to install.

How do I get it running?

The commands in the “Quick start” box above come straight from the README. Paste them into a terminal; the README covers options and troubleshooting.

Where does my data go?

T h

What doesn't it do yet?

A

Is it finished?

Not yet: it's marked “live” and still changing. The README always has the current state.

Is it free?

The source is public on GitHub under the MIT licence.

Who made it?

Juwen Lie, an R&D mechanical engineer in Batam, Indonesia, as an independent personal project.

Keep exploring

More from the same workbench.

Scan a skill now.

Check an AI agent skill or MCP server before you trust it