DocsUse itLeaving an agent running

Leaving an agent running

The point of Memnox is to let an agent work while you are not watching. Four steps get you there.

1. Ask whether it is ready

bash
memnox unattended --agent claude-code
If claude-code runs on its own
on its own     0
waits for you  7
never          7
no rule        114
 
Settle these first
   114 capabilities have no rule at all, more than the 14 that do
   2 calls are waiting on you
   4 sessions are paused by the breaker

Work through the list it prints. When it says nothing needs you, the agent can be left alone.

2. Stop being asked what you always approve

bash
memnox next --agent claude-code   # what it could do alone, action by action
memnox next --hand-over           # write allow rules for everything that is ready

next reads what you already approved and never refused. One refusal keeps an action out, whatever the count.

3. Tell it the job

Declare what you asked for, so anything outside it stands out:

bash
memnox run --task "fix checkout" --paths 'src/checkout/**' --expect 40 -- claude
FlagMeans
--taskWhat you asked for, in your words
--pathsThe files the task covers
--expectRoughly how many actions it should take
--investigateRead only: nothing outside this machine changes

Or set it for the repository, so every agent there sees it:

bash
memnox task set "fix checkout" --paths 'src/checkout/**' --hours 4
memnox task show
memnox task clear

4. Cap what it can spend

bash
memnox budget suggest                                          # a generous starting set
memnox budget set deploys --actions 'vercel.*,railway.*' --limit 5 --window day
memnox budget                                                  # what is left

A budget counts calls, or dollars with --unit usd, per day, hour or session.

While you are away

  • A call that needs you waits, and reaches your Mac or iPhone.
  • A loop that fails the same way five times is paused, so it stops spending. memnox resume <session> --by <you> lets it carry on.
  • When you are back:
bash
memnox report --since 1d     # what agents did, and what was redone
memnox claims                # what the agent said against what it did
memnox rewind                # if you do not like what you see