Sunicorn · Control and security
Held to the standard of finance.
A Crew works under rules you set, and every action it takes is recorded. This page lists what holds it: control, the record, security and data.
01Control
Each kind of action runs alone, after a veto window, or only with your yes.
First
It prepares
you decide each move
Then
It acts
in your name, where you can stop it
Then
It runs
a line of work alone, within limits you set
A kind of action moves up only as its record earns it, and only a person raises it.
- Rules in code. The agent cannot change them. Anything that leaves goes first through a sender that checks them.
- One stop switch. It halts the agents and everything acting for them.
- Two kinds of interface. Only your group and approved addresses reach the agent. What others see is computed by code.
02The record
Every action is recorded, out of the agent’s reach. The monthly report sets the agreed outcomes against the fee.
See every Crew at work, hour by hour →
Counts only, no names.
03Security and data
- A machine reserved for you, in the hosting region of your choice, Frankfurt by default, with its own data and keys.
- Administration only over a private encrypted network.
- The model comes from a US provider today, with training off. A European provider is possible.
- Every night, an encrypted copy is stored in the EU.
- Every offer comes with a data processing agreement that names every sub-processor, and an NDA.
04Common questions
- Mistakes: Sunicorn tells you, then fixes it. Decisions stay yours.
- Your clients: nothing leaves before it has been shown to you, with time to stop it.
- Your team: the Crew takes the volume and the follow-up. Judgement and relationships stay with people.
- Leaving: one month’s notice after the minimum term. Your data comes back within 30 days in open formats, then is deleted.
05How it is built For technical readers
By role, without versions or network details.
- One machine per client, with its own agents, memory, data and keys. Nothing is shared between Crews.
- The agent proposes and code decides. Mail and payments leave through a separate sender that holds the credentials and checks the rules first: allowlists, caps, cadence, veto windows and approvals.
- The rules are versioned in code the agent cannot change. A new version is recorded before anything runs under it.
- One stop switch halts the agent, its schedule and every service acting for it.
- Every turn, tool call, model request, approval and effect is one line in a chained, append-only record, out of the agent’s reach. It holds identifiers, numbers and fingerprints, never content, and is copied off the machine daily.
- Administration runs only over a private encrypted network. The machine has no public admin port.
- Every night, an encrypted copy goes to off-site storage in the EU under a 30-day deletion lock. A sample is read back each week.
- The model is set by configuration: a US provider today, with training off on the account.
- Interfaces serve people and their own agents through standard protocols (MCP and A2A).
- Each machine’s health is checked every 15 minutes. The counts on the Live page come from these records.
A short mail: your firm, the work you would hand over, who introduced you, and two good times.