Ask about the method.
The beta interface: the hosted chat, talking only. It is bring your own key — this site holds no model credential and pays for nobody, so open Settings, paste a key, and it pays for your own turns. If an answer proposes a tool call, the answer says so and runs nothing.
Asking this backend what it can run…
Continue on a machine you own
Nothing in this tab can run a capability: this deployment has no
scope file, no dated attestation, no audit log and no process to
start. The same page does run, on a machine that has all four -
red serve --hands adds that surface and nothing here
changes it.
red serve --hands
It serves this same file on loopback - on the address its own banner
prints, 127.0.0.1:8000 by default - with the run bar
answering for real: every plan is the guard's answer about one exact
command line, and every run is written to the audit log before its
process starts. The download installs the same agent this page is
talking about, so the conversation continues there, under the scope
you write.
Settings
Where your turns go, and how hard each one may work. Sent with every message, stored nowhere.
Bring your own key. This site keeps no key of its own. Yours is used for one question and stored nowhere. No key, no answer.
Reasoning power — how much thinking one answer may spend, and how long it may be.
Execution power — nothing here can run, so this sizes one turn only: how much chat it carries, and how long the site waits. It never buys you more turns.
Enter sends, Shift + Enter adds a line. Five turns per visitor per ten minutes, all on your own key. Nothing here runs a scan — that is the CLI, behind scope and an authorisation.