An independent check from a different company
The AI that builds a change never reviews it. A model from a different company reviews the exact commit, and its verdict is bound to the bytes it read. A review of anything else is refused.
Biclave exists to control what AI agents do. These are the controls it applies to every change today, and the way we hold Biclave itself to the same standard.
The AI that builds a change never reviews it. A model from a different company reviews the exact commit, and its verdict is bound to the bytes it read. A review of anything else is refused.
Your rules live in a written charter that you approve by its SHA-256 hash, together with the runner's own code and prompt templates. Change any of them and Biclave stops acting on its own until you approve again.
Every decision carries a fingerprint of the question, its options and its context. An answer applies only to the decision it was given for, you confirm it with a typed phrase, and acceptances keep a veto window open.
Agent steps run inside the operating system's sandbox behind a command guard, with writes limited to the work in hand. Review runs have the network turned off. Credentials live in the operating system's keychain and never reach the logs.
Every transition, decision, notification and recovery is written to an append-only log before anything acts on it. The log is chained and sealed, and Biclave checks the seal itself.
Stop rules, daily budgets and a bounded shutdown keep a run inside its limits. When something does not add up, Biclave asks a person instead of guessing.
Today Biclave runs on your own infrastructure. Your code, your evidence and your decisions stay on your machines. Code is sent only to the AI providers you choose to use, under your own agreements with them.
Every change to Biclave itself goes through the same two keys: it is reviewed by an independent AI from a different company, and it is tested by more than 3,000 automated tests before it is accepted.
If you believe you have found a security issue in Biclave, email security@biclave.com. Please give us a reasonable time to fix it before you share it publicly. We will acknowledge your report and keep you informed.
Questions about security for your team? Talk to us.