I built a way to leave Claude Code working overnight
Claude Code kept stopping on one small "Allow?" while I slept. So I designed Away, a free plugin that keeps it working and leaves me a short list in the morning, and learned what it takes to trust an AI you're not watching.
Purvang Mehta · Product designer
28 September 2026 · 3 min read
I used to give Claude Code a big job before bed, and wake up to nothing done. It had stopped at 1am on one small "Allow?" and waited all night.
The fix turned out to be a design problem, not a settings problem.
4
design decisions
3
levels to choose from
100+
automatic checks
3
systems: Mac, Windows, Linux
Why full permissions weren't enough
Claude Code has a mode that skips most permission prompts. I had it on. It still stopped: on questions, on plan approvals, on new websites, and whenever it decided it was done.
Full permissions alone
- Stops on questions
- Stops when it thinks it's done
- Deletes and deploys without asking
- Morning: scroll everything
With Away
- Questions become noted decisions
- Sent back until the job's done
- Risky actions wait for you
- Morning: one short list
So removing the stalls also removed every safety check.
The real question was trust
Permission prompts are how an AI checks with you before something you might regret. Removing all of them isn't a feature. It's a gamble.
Every AI product that works in the background runs into this question.
Four decisions
Away is a small plugin: /away:on and the job before you leave, /away:off when you're back.
Decide what waits
Normal work goes ahead. Anything irreversible waits for you.
You hold the off switch
Only you can switch it off. Never the AI.
Decide, and say so
No questions at 3am. It picks, and notes why.
One list in the morning
Done, skipped, and decided for you. Nothing else.
Here's a night, one action at a time:
- 23:10Run the test suiteApproved
- 23:14Edit 12 image filesApproved
- 00:02Delete the old build folderNeeds you
- 01:37Ask: 16:9 or 4:3?Went with 16:9, noted
- 02:15Push to mainNeeds you
- 03:40Stop and ask to continueSent back to work
- 05:05Job doneAway switched off
- 07:30Your morning list2 need you
A moon in the Mac menu bar fills while Away is on, and shows a number when things wait for you.
Testing it like it could hurt me
I tried to break it on purpose, in a live session. It failed in useful ways:
- Question cards still appeared. Now they're blocked.
- The browser's site pop-up waited forever. Now Claude stays out of the browser at night.
- With no prompts, risky actions ran unseen. Now every action is checked first.
- Claude could switch Away off. Now only you can.
Today it has over 100 automatic checks, passing on Mac, Windows and Linux.
What it isn't
Away reads each command before it runs. It can't see inside a program Claude writes and runs. It's a seatbelt, not a locked room. For real isolation, pair it with Claude Code's built-in sandbox. A tool that promises total safety gets trusted too much.
For teams building AI products
Autonomy is a dial
Let people choose how much it does alone.
The person keeps the off switch
The AI never removes its own limits.
Decide and disclose
A reported choice beats a blocked question.
Summarise honestly
One short list beats a long log.
Say what it can't do
Trust comes from limits people can see.
Try it
Away is free and open. In Claude Code:
/plugin marketplace add thepurvangmehta/away
/plugin install away@awayThe code and its known limits are on GitHub. Still finding where it breaks. If something stops it anyway, tell me.
