We build AI that
knows its limits.

MeTal LLC is an AI company based in California. We make AI products and infrastructure that take safety seriously — not as a marketing line, but as a technical commitment.

A small team building something that matters.

MeTal started with one idea: that AI being useful and AI being safe aren't in conflict. They're the same goal, approached from different angles.

We're a small company. We don't have a hundred engineers or a billion-dollar compute budget. What we do have is a clear set of values about how AI should behave — and we hold ourselves to them.

We build Aurora, our AI system, and MeTal Cloud, a hosting platform for developers. Everything we ship is designed to be transparent, predictable, and honest about what it can and can't do.

Constitutional AI as a foundation.

We follow Anthropic's Constitutional AI framework — a written set of principles that AI systems are evaluated and trained against. Not just filtered outputs, but values baked into the model itself.

This means Aurora isn't just trained to be helpful. It's trained to be honest when it's uncertain, to decline requests that could cause harm, and to support human oversight rather than undermine it.

We publish our principles openly. If our AI does something that violates them, we want to know — and we'll fix it.


What we stand for.

Safety over capability

We will always choose a safer, less capable system over a more powerful one that poses unacceptable risks. Capability is meaningless if it causes harm.

Transparency

We're open about what our AI can do, what it can't do, and how it makes decisions. No black boxes. No marketing spin.

Human oversight

AI should amplify human judgment, not replace it. Every system we build keeps humans in the loop and makes it easy to correct mistakes.

Accountability

When something we build causes harm, that's on us. We take responsibility for the systems we ship and maintain them accordingly.

Built on Constitutional AI

We follow Anthropic's Constitutional AI research — a framework for training AI systems with explicit written values, not just output filters. Aurora is evaluated against these principles continuously.

Read the Research