Built for the
unattended hour.
Neku came out of one failure mode: an agent runs for an hour, stops on a command it will not run alone, and nobody is watching the terminal it stopped in.
Why we built Neku
We ran long agent sessions on machines in the next room. The same shape of problem kept showing up: the agent had been idle for forty minutes, not stuck, just holding a tool call and waiting for a person.
The workarounds were worse than the problem. A terminal on a phone is not supervision. A notification with no command in it is not a decision. A policy file that allows everything is an unattended agent with paperwork.
What Neku is not
- — Not a terminal emulator.
- — Not a cloud runner for your agents.
- — Not an agent that acts for you.
- — Not a log aggregator.
- — Not a place your code is stored.
supervision layer only · your hosts run the work
Three rules we do not bend
Fail closed, always.
A held call is denied when its countdown runs out. Timeouts deny, the agent is told why, and the denial is logged. Auto-approve-on-timeout does not exist.
Your machines do the work.
Code, transcripts, diffs and terminal output stay on your hosts. The relay carries only sealed bytes it cannot open, and Neku never copies a transcript into its own store.
A control surface, not a terminal.
Neku shows state, spans and held tool calls and carries your decision back. It does not stream a pty and you cannot type into the agent through it.
Who is building it
Daemon, transport and host adapters.
Mobile and web apps, gestures and the inbox.
Policy engine, approvals and the parser contract.
We are hiring carefully
Systems, daemon and mobile. Open roles and headcount: [ROLES] — write to us and say what you have shipped.