Security
One gate,
no exceptions
Agents are useful exactly to the degree you can trust the walls. So the walls came first: every read and every action an agent takes passes through a single enforced gate — in code, not in a prompt.
The ask
An agent wants something: read a room, open a file, send a message. Every ask is explicit — nothing is ambient.
The broker
One gate checks the ask against the agent's badge — the same badge system that scopes what your teammates see. Not a system prompt asking nicely; a boundary the model can't talk its way past.
The answer
Allowed, held for a human's approval, or denied. Whatever happens, it's written to a log in plain language.
Your keys stay yours
Agents run on your own Claude account. Budgets, caps, and the relationship with Anthropic stay in your hands — we never resell tokens.
Approval holds
Anything that leaves the building — email, invoices, messages to clients — can be held until a person says yes. You choose where the line sits.
A log you can read
Every allowed, held, and denied ask is recorded in plain language — who asked, for what, and why the gate answered the way it did.
Your room, walled
Each greatroom is isolated from every other. Your data isn't training material, and you can export the whole room whenever you like.
In practice
What the gate
looks like
Three real answers from a working room. The denial is the point — a permission system you never see saying no isn't one.
Where does my data live?
In your greatroom, encrypted in transit, isolated per company. Docs you drop in belong to the room; delete the room and they go with it.
Certifications?
We're a small team in private beta. SOC 2 is on the roadmap, and we won't claim it a day before it's true. Until then, this page says exactly what we do and don't do.
Found something?
Tell us at security@greatroom.ai. We read every report, respond fast, and credit researchers who help us make the walls stronger.
Trust is the product. Come test it.