// The guide
Claude Code permissions: auto mode is bypass with a leash.
A coach we onboarded ran Claude Code with all permissions bypassed. Their reason was honest: "saying yes 100 times a day is unbearable." They were right about the problem and wrong about the fix. The fix is auto mode — nearly all the silence of bypass, with a leash that still catches the actions you would actually want to review.
This comes from real onboarding calls, July 2026.
The three modes, honestly described
- Manual (the default). Claude Code asks before every action. Safe, and exhausting. This is what drives people to bypass in the first place: the hundredth "yes" of the day teaches you to stop reading what you are approving — which quietly defeats the point of asking.
- Auto. An evaluation layer classifies each action. Safe actions run on their own; risky or destructive ones still stop and ask. You go from a hundred prompts a day to a handful.
- bypassPermissions. Nothing asks, ever. With today's agentic sessions — long-running, multi-step, touching real files and real services — this is genuinely dangerous. One misread instruction and there is no checkpoint between the model and the consequence.
What auto feels like in practice
In our experience roughly 99% of permission prompts simply disappear. And the rare stop is exactly the one you want: auto mode has blocked real mistakes for us — an unreviewed overwrite of a live document, and an unapproved background job. Both times, the stop was the system working, not the system nagging.
That is the trade in one sentence: bypass removes the noise and the brakes; auto removes the noise and keeps the brakes.
Switching takes ten seconds
In Claude Code:
/config → permissions → default mode → autoThat is the whole migration.
In defense of the coach
One honest note. Auto mode is newer than bypass, and plenty of people simply do not know it exists. Before it shipped, bypass was the only escape from prompt fatigue, and choosing it was a rational answer to a real cost. The coach was not reckless — they were solving yesterday's problem with yesterday's tool. If that is you, nothing is wrong with your judgment. The better option just arrived quietly.
The pattern behind this post is one we see constantly onboarding people onto Klyr: the setup someone runs is usually the best answer to the tools they knew about at the time. Part of what Klyr's onboarding does is walk through exactly these settings — so your assistant runs fast without running unsupervised.
New to Claude Code? Start with our non-developer guide.
FAQ
What permission modes does Claude Code have?
Three that matter here: manual (the default — asks before every action), auto (an evaluation layer runs safe actions alone and stops only for risky or destructive ones), and bypassPermissions (never asks — dangerous with agentic sessions).
Is bypassPermissions safe?
Not with modern agentic sessions. Nothing stands between a misread instruction and a destructive action. It made sense before auto mode existed; today auto gives you nearly the same silence with the dangerous stops kept.
How much does auto mode actually reduce prompts?
In our experience about 99% of prompts disappear. The remaining stops are the ones worth reading — for us they have included an unreviewed overwrite of a live document and an unapproved background job.
How do I switch Claude Code to auto mode?
Run /config, open permissions, set default mode to auto. Ten seconds.
Does auto mode still ask before destructive actions?
Yes. That is its point: safe actions run without asking, risky or destructive ones still stop for approval.