About Us

The security layer for AI agents

We build open-source security tools for AI agent stacks. Research-backed. Community-driven. Always updating.

Our Philosophy

Why agent security matters

AI agent security is an open problem. We're solving it in the open. OpenClaw skills, MCP servers, and agent plugins run with access to your credentials, file system, and network, yet most developers install them without a second look.

We found that 7.1% of skills on a major marketplace were actively stealing credentials or exfiltrating data. Tool poisoning, prompt injection, and credential harvesting are real threats, and nobody was scanning for them.

Firmis exists to close that gap. We scan your entire agent stack, flag threats in plain English, and give you actionable fix instructions. Education first, not fear first.

Our Values

What we believe

Open Source

Apache-2.0 scanner, ELv2 engine. Inspect every rule, run offline, contribute back.

Research-First

Every detection rule is backed by data. We analyzed thousands of AI agent skills to build the threat model. Data over opinions. Always.

Developer-Friendly

CLI-native, zero friction. One command, plain English results, no account required.

Community-Driven

Every scan contributes signal. New threats found by the community become detection rules within hours. Everyone who scans makes everyone safer.

Our Team

Who we are

Founded by engineers who saw the agentic security gap firsthand. As AI agents proliferated across development workflows, nobody was auditing what these tools could actually access.

We built Firmis to give every developer the same visibility into their agent stack that enterprise security teams have. For free, with one command, in plain English.

How we build

Experiment-based

Every feature ships with a hypothesis and a way to measure it. We run experiments, not roadmaps. If the data says we were wrong, we change course.

Research-backed

We scanned thousands of agent skills before writing our first detection rule. Every claim we make is traceable to data. We publish our methodology so others can verify.

Community-driven

The scanner is Apache-2.0. Every scan contributes anonymous signal to the shared threat feed. New threats become detection rules within hours, not weeks.

Education first

We explain threats in plain English, not fear-driven marketing. Every finding tells you what happened, why it matters, and exactly how to fix it.

Get Started

Scan your agent stack

One command checks any AI agent platform for malicious tools, exposed credentials, and data theft. Free. 30 seconds. Plain English results.