The five components
Five named components.
Each with one guarantee.
What each does. What it guarantees. What you see. How it works inside is the part we keep to ourselves. That's the next decade.
Substrate
Persistent memory.
What it learns, it keeps. Across every session, every restart, every model swap. There is no "start over" mode because there is no "forgot you" mode.
What it learned yesterday answers your question this morning.
Workers
Autonomous background workers.
You set a goal. They pick it up. They run on cheap models when cheap is enough and on smart ones when smart is necessary. They sleep when there is nothing to do, and when they sleep you pay nothing.
Set a goal at midnight. Wake up to the work done. No paid context window ticking while it waits.
Dispatchers
Per-question routing.
Every request goes to the cheapest brain that can answer it correctly. You don't pay for genius when reflex will do. You don't get a reflex answer to a question that needs depth.
Cheap models for cheap work. Smart ones when they matter.
Operator Vault
Cryptographic credential isolation.
Your keys, tokens, secrets are sealed by your operator key. The model receives outcomes, what an API returned, never the credential itself. Even we cannot read what you've sealed.
A prompt that asks for your keys gets nothing. They were never readable
Sealed Body
Signed, isolated runtime.
The mind lives in its own machine on your laptop. Signed at boot, isolated from your host, with a kill switch you control. We don't have a back door. There is nothing to phone home.
You can pull the plug. It can't escape.
This is what it does. How it does it is the part we don't publish. Read /protocol for what it says, /credibility for what we measure.