One business.
One sealed instance.
An instance is everything your business's AI needs, built only for you: the assistant your team talks to, the agents that do the work, the gate that checks every call, your data, an auditor and a shared workspace. Here is what is inside and how a request moves through it.
Six parts, none shared with another customer.
Client Operating Agent
The only agent your people talk to, on the web dashboard or Telegram. It keeps each person's memory and files separate, routes work, and checks every answer before it's shown.
Agents
Specialists that do one job each, with read-only tools for the records they're allowed to see. Each runs in its own container, as its own non-root user.
Identity gate
Every call between agents carries a ticket that is valid once, for 60 seconds, for one caller and one route. The gate records who asked, what for, and the decision.
Your data
Each agent writes only its own records, and only that agent can write them. What you author is versioned and backed up before every release.
Guardian auditor
A read-only auditor inside your instance watches logs and changes. It can raise an alarm. It can't change anything.
Shared workspace
Projects, tasks and files for your team, signed in through the same dashboard, with owner and staff roles.
Your instance reaches only two things beyond itself: private AI models over a sealed link that admits nothing else, and The Foundry, which delivers approved releases. It has no route to any other customer.
"Is anything overdue?" takes seven checks to answer.
The sequence is the product. Each step either passes, or stops and tells you why.
A signed-in person asks
The dashboard knows who you are. Your question is filed under your own identity, never shared with a colleague's.
The assistant picks the agent
It routes only to agents your instance actually has installed. If none fits, it says so.
The call gets a ticket
The gate issues a single-use, 60-second ticket naming the caller and the route.
The agent checks the ticket
No valid ticket, no answer. Every decision is written to the gate's log.
The agent reads its own records
Through read-only tools, and it returns evidence, not prose.
The assistant drafts once, then checks the draft
Every figure, date and "you told me" must rest on the evidence or your own saved words. Claims of actions it didn't take are refused.
You get one of four outcomes
Grounded, couldn't answer, couldn't look, or held back. Never an unchecked guess.
Nothing goes live by accident.
Your instance runs from a pinned release. A change merged in our code does nothing to you until a release is approved and moved.
Built on a throwaway copy
Written and tested on an instance that is created per change and then destroyed.
Accepted on a standing copy
Proven end to end on a test instance with realistic data before anyone approves it.
Your data is snapshotted
Before every release, and re-checked after, byte for byte.
Moved to your pinned copy
Only with a named approval. A release can't write your data.
Checked live
Every agent's health and the gate itself are checked after the move. A failed check says so.
Web dashboard
Your assistant, documents, workspace and full conversation history, searchable by date or word.
Telegram
The same assistant and the same memory on your phone, linked to you with a one-time code from your signed-in dashboard.
Owner controls
Invite staff, set roles, and see which agents are working for you and whether each is healthy.