The 4 Modes in Claude Code
One shortcut, four gears — from “ask me everything” to “go and build it.” Which to use, when, and the bits people get wrong.
What Just Changed
Auto is now the mode you start in. Open Claude Code and it is already in auto, where it gets on with the work and a safety classifier vets each action rather than you approving them one at a time. Enough people were switching to auto by hand that it became the sensible place to begin.
What you used to start in is now called manual. Same mode, asking before every edit and every command, just renamed to something that describes it now that it is no longer the default. You can still get to it, and on a sensitive job you should.
Auto is not switched on everywhere. It needs a recent Claude Code and it is enabled per session, so if you cycle through and never see it, that is why.
The real consequence is the one nobody is talking about. A mode that asks permission constantly is also a mode that keeps interrupting you into noticing what it is about to do. Take that away and a vague instruction goes a very long way before you see it. So the answer to auto being on by default is not to turn it off. It is to spend longer in plan mode first.
Most people use Claude Code in one mode the whole time and never find out there are four. You flip between them by pressing Shift + Tab, and each one changes how much Claude asks versus how much it just gets on and does. What changed recently is which one you land in when you open it. Here's what each is for.
The Four Modes
Manual
Ask before editsIt asks before it does anything — every edit, every command. Nothing happens without your say-so.
Use it: When you're new to it, or working on something sensitive you want to watch closely. This used to be where every session started, and it is now something you choose.
The trade
Safest, but slowest. If you're approving the same kind of thing over and over, that's your sign to move up a gear. Claude Code calls this one both “manual” and “default” — same mode, two names, left over from when it was the one you started in.
Accept edits
Edit automaticallyIt stops asking before file edits — and a handful of safe file commands like making folders or moving files — and just makes the changes.
Use it: When you're iterating on code and happy to review everything after in one go (in the git diff, say) rather than approving each change inline.
The honest bit
It does NOT say yes to everything. It still stops and asks before it installs a package, pushes to GitHub, or goes on the internet — and it won't touch protected files (.git, .claude) without asking.
Plan mode
Plan modeRead-only. It can't edit, create, or run anything that changes your files — it reads, explores, and hands you a plan first. Approve the plan and it switches into a working mode and starts.
Use it: When you really want to plan out a bigger build, change or feature before it touches a thing.
Good to know
It doesn't lock you into the first plan — re-plan as many times as you like before approving. Shift+Tab again to leave plan mode without making any changes.
Auto
Auto modeNow the defaultIt does almost everything itself — but a separate safety AI reviews each action before it runs and blocks the genuinely dangerous stuff (piping the internet into your shell, force pushes, production deploys, mass deletions, hard resets).
Use it: Long, multi-step jobs you trust the direction on. This is where most of the time goes — the models are good enough now that if you spec it well up front, it delivers almost exactly what you asked for. It is also where you now start, unless you change it.
The honest bit
It's not no-guardrails — but it's not magic either. Needs a recent Claude Code and model. Don't point it at anything you genuinely can't undo. If the safety AI blocks the same move a few times, it pauses and hands back to you.
How to Tell Which Mode You're In
In VS Code: the mode indicator sits at the bottom of the prompt box — a little icon tells you where you are. Click it to switch.
In the terminal: the status line at the bottom shows the chevrons and the mode name (⏵ default, ⏵⏵ accept edits on, ⏵ plan mode, ⏵⏵ auto mode on).
What People Get Wrong
✗ The myth
Accept edits means it says yes to everything.
The truth
No — installs, git pushes and network requests still stop and ask. It only auto-approves edits and a few safe file commands.
✗ The myth
Auto mode is the same as having no guardrails.
The truth
No — a separate safety AI still reviews every action and blocks the risky ones. The no-guardrails mode is a different thing entirely (below).
✗ The myth
Plan mode commits you to the first plan it writes.
The truth
No — you can refine and re-plan as many times as you want before you approve anything.
The one to leave alone
Bypass permissions
There's a fifth setting that turns every safety check off — no prompts, no classifier, nothing. You can launch straight into it with a deliberate flag (--dangerously-skip-permissions), and where it's enabled for your session it now sits in the Shift + Tab cycle too, between plan and auto. That makes it easier to land in by accident than it used to be, so know what it looks like. Only safe inside an isolated container or VM. On your normal machine, leave it well alone.
Four modes, one shortcut.
Now that auto is the starting point, the mode worth more of your time is the other one. Plan a bigger build in plan mode, let it keep asking questions until you're genuinely aligned, then switch to auto and let it run. Measure twice, cut once. Point auto at a sloppy prompt and it will build you something sloppy, faster than ever.
Want to get better at this?
Let's set up your agent ecosystem.
This page is one job. The whole thing is a set of them running off the same foundation, the folder your AI works out of, the brief on how you like things done, the accounts it can reach and the skills that run off all of it, so your admin gets done the way you would do it. It is what I teach one to one, on your own machine and against your own work, and the quickest way to find out what yours would look like is a proper chat about it.
Not ready for a call? Start with the free agent series, what an AI agent actually is, and build up from there.