Every useful AI agent eventually asks for something private. A passport number to book a flight. A policy number to file a claim. An invoice to reconcile your books.
Today, there are two bad options. You paste the data into a chat and lose track of it forever. Or you connect the agent to your whole Drive and hope it only looks where it should.
We think there's a third way: give permission, not possession.
Ask. An agent asks your vault for something, through MCP or our permissioned API. It never sees the vault itself, only the answer.
Check. Your rules decide, and they're enforced by code, not by an AI. A model can't be talked into breaking a rule, because a model isn't the one deciding.
Release the minimum. Often the answer can be just "yes": the passport is valid, the policy is active. Hushkey signs that answer and no data leaves. When the agent truly needs a field, it gets only that field, only for this task, and the access expires.
Record everything. Every request is signed, timestamped and logged. You can see who asked, for what, and when, and you can revoke any agent in one tap.
The result is agents that are genuinely useful and a life that stays yours.
Your agents get answers. They never get the vault.