Tool · CLI
ads
An agent team tests, changes and reviews each change. You keep the last word.
You ask Claude Code for a change, “fix #12” or “work the queue”, and ads
runs it for you. A test author writes tests first and freezes them (a
docs-only change has none), an implementer makes the change until those tests and the checks pass, and
separate reviewers look at it in layers. The more risk, the more layers.
The risk of the change decides who signs it. Once ads init has set up the
team key, a trivial, low or medium change within its size limit that passes
every check lands by itself, signed with that key; a medium change with a
dispute or a stall waits for you. A high or critical
change waits for your sign-off: one block in the session with the facts of
the change, and a short ok <code> from you, which signs it with your key.
Every step is a signed event in git, and every landed change can be undone
with a signed revert.
Each change has a budget of rounds, seconds and tokens for its risk, and asks
before it spends more. You can pause, stop or undo a change by saying so. The
approval guard is a text filter that runs as your user, not an OS sandbox,
and ads needs Claude Code. It is in beta, and the repository is private
for now.
