Hi Product Hunt, I'm Faisal, the maker of PhysicsBase.

AI agents are starting to touch real systems: refunds in Stripe, records in a CRM, emails to customers. The scary part isn't the model. It's that one bad instruction, from a hallucination or a sentence hidden in a web page, can refund, delete or leak, and most teams only find out later.

PhysicsBase sits between your agents and those systems:

Your keys stay in a vault. Agents never see them, so there's nothing in their context to steal.

Rules on every call. Refunds up to $200, never a card number, never the same refund twice.

Only the irreversible waits for you. Payouts and emails pause with a plain-English preview; everything else just runs.

Undo. Before an agent changes or deletes a record, PhysicsBase keeps a copy. Undo one change, or rewind the last hour.

It's free to start (paste one URL into Claude, Cursor or any MCP client), open source under Apache 2.0, and you can try to break a live agent on the homepage without signing up.

I'd love to know: what would you need to see before letting an agent touch production? I'm here all day to answer anything.