You're deep in a refactor, Claude needs to edit a file, and you hit Enter to approve. Then it needs to run a test. Enter. Then another edit. Enter. Fifteen approvals later, you realize you haven't actually read any of them.
The alternative is --dangerously-skip-permissions, which lets Claude do anything without asking. That's... not great either.
Auto mode is the middle ground. A classifier evaluates each action: safe edits and commands run without interrupting you, anything destructive gets blocked and surfaced.
Key Takeaways
- Auto mode uses a classifier to approve or block actions automatically
- Safe operations (file edits, reads, standard commands) proceed without prompts
- Destructive operations (force push, rm -rf, production deploys) get blocked
- You can cycle between permission modes with Shift+Tab
- It's a research preview — the classifier improves over time
How It Works
When Claude wants to use a tool, the classifier evaluates the action against a safety model. The classifier considers what tool is being used, what parameters it's receiving, and the current context.
Approved automatically: file reads, file edits in your project directory, running tests, git status/log/diff, standard development commands.
Blocked and surfaced: rm -rf, git push --force, commands targeting production, anything outside your project directory, operations that could cause data loss.
The key insight: you're not giving up control. You're delegating the boring approvals (yes, Claude, you can read that file) while keeping oversight on the ones that matter.
Getting Started
Quick toggle during a session: Press Shift+Tab to cycle through permission modes. You'll see the mode indicator at the bottom of your terminal change.
Set as default: Add this to your project or user settings:
{ "permissions": { "defaultMode": "auto" } }
This goes in .claude/settings.json (project) or ~/.claude/settings.json (global).
When Auto Mode Shines
Long coding sessions. A 2-hour refactor might involve hundreds of file operations. Auto mode handles the routine ones so you can focus on reviewing the actual code changes.
Multi-file edits with tests. Claude edits a file, runs the test, sees the failure, edits again. Without auto mode, that's four approval prompts for one iteration cycle.
Exploratory work. When Claude is reading through your codebase to understand a problem, every file read triggers a prompt. Auto mode lets it explore freely.
What Won't Change Your Mind
If you're uncomfortable with any automated approval, auto mode isn't for you — and that's fine. The standard mode with explicit approval for each action is still the default.
If you're already using --dangerously-skip-permissions and trust Claude completely, auto mode adds safety checks you might not want.
Auto mode is specifically for people who want fewer interruptions but aren't willing to skip all safety checks.
The Honest Limitations
⚠️ It's a research preview. The classifier will occasionally block something safe or approve something you'd rather review. Report these cases — the classifier learns from feedback.
❌ Auto mode doesn't learn your personal preferences yet. It uses a general safety model, not one tuned to your project's risk profile.
❌ It can't distinguish between your staging and production environment by name. Acurltoapi.example.comlooks the same whether that's staging or prod. Use hooks if you need environment-aware blocking.
Combining with Hooks
Here's the thing most people miss: auto mode and hooks complement each other perfectly.
Auto mode handles the general "is this safe?" question. Hooks handle project-specific rules. For example:
- Auto mode approves a
git commit(safe in general) - Your PreToolUse hook checks that the commit message follows your convention (project-specific)
Auto mode is the safety net. Hooks are the guardrails.
Setting Up
Auto mode is built into Claude Code v2.1.83+. No installation needed.
- Update Claude Code to the latest version
- Press
Shift+Tabto toggle, or setdefaultMode: "auto"in settings - Watch the mode indicator at the bottom of your terminal
Full details in the Permission modes guide.