An agent platform where
nothing runs unwatched
Neutron is one program that many separate teams can run, each with their own AI agents. Its reason to exist is the part most agent tools skip: deciding, in advance and in writing, what an agent is allowed to do — and stopping it when it tries to do more.
Governed autonomy
Every mutating action passes a gate before it runs. Three outcomes, not two: allowed, parked for a human, refused — and all three are recorded.
One image, many customers
A customer is a configuration file and a database, never a fork or a branch. Isolation is physical: its own pod, its own data.
You cannot approve yourself
Separation of duties is written into the permission policy as a subtraction, so the engine enforces it rather than etiquette.
Interchangeable engines
Codex, the Claude Agent SDK, and OpenCode sit behind one seam. No provider type is allowed to escape it.
Code without shipping rights
An issue becomes a reviewable merge request. The pipeline never merges and never deploys — autonomy ends at the merge button.
A swarm, not a server
A hub instance drives peer instances and a desktop node over the same authenticated interface.