MAGI SYSTEMS
Open Source AI Security Tools
About
Magi Systems builds open source tools for AI that has to be trusted. Everything here runs locally, carries no dependencies, and shows its work: the source is public, so you can read exactly what it does before you run it. Take them as they are, or we will shape them around how your organization actually operates.
Every tool we publish runs on your own machine. No shared infrastructure, no telemetry, no calling home. Several of them make no network requests at all, so they work behind an air gap exactly as they work on a laptop.
The tools speak the language auditors and regulators actually cite: CIS Controls, ISO/IEC 27001 and 42001, NIST AI RMF and CSF 2.0, SP 800-53, the SOC 2 Trust Services Criteria, and CCPA/CPRA. A control id that does not exist fails validation instead of passing as real.
These are the tools we use in our own work, published under MIT rather than rebuilt as marketing artifacts. Read the source, run the tests, fork it. Nothing is gated behind a form. When the default shape does not match your environment, we adapt it to fit.
The thread running through both tools is refusing to take a claim on faith. Evidence defaults to unverified, grades are a deterministic function of the findings, and an AI assistant has to show tool evidence before it calls work done. Confident prose is not proof.
Open Source
The evidence engine for security and AI governance. It grades a risk assessment as a deterministic function of the findings. It cannot skip the grade or invent a threshold, and evidence defaults to UNVERIFIED. Eight bundled control catalogs, zero dependencies, no network. MIT.
Learn More →Make any AI coding assistant verify its work before it says "done." A plain-markdown operating doctrine you drop into a project: hard rules, a six-phase algorithm, and a verification doctrine that refuses unverified claims. Works across Claude Code, Cursor, Codex, Aider, and Copilot. MIT.
Learn More →Everything is public on GitHub: read the source, run the tests, open an issue, or fork it.
Make It Yours
Clone the repository and run it today. The tools are complete, tested, and documented, not trial versions with the useful parts removed. Most organizations never need anything beyond what is already published.
No two environments grade risk, ship code, or answer to auditors the same way. We adapt these tools to your control catalogs, your pipelines and ticketing, your policies and approval gates, so the output matches what your reviewers actually expect to see.
Tell us what your environment looks like and we will tell you honestly whether the stock tool already covers it. Start that conversation →
Get In Touch
Questions about a tool, a bug you hit, or something you want to contribute, send it here and it lands straight in our inbox. Bug reports and feature requests are better off as a GitHub issue, where everyone can see them.