What is auto mode in Claude Code?

Auto mode is a Claude Code permission mode where a separate classifier reviews every tool call before it runs instead of prompting you for each one. It allows routine actions and blocks ones that are irreversible, destructive, or aimed outside your environment. If Claude is blocked three times in a row, or twenty times in a session, Claude Code falls back to manual approvals.

When does auto mode become the default in Claude Code?

Starting August 14, 2026, new Claude Code sessions on the Pro, Max, and Team plans run in auto mode by default. If you already pinned your own default permission mode, nothing changes. Claude Enterprise, the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry remain opt-in for now, with Anthropic saying it plans to make auto mode the default there in the following month.

Does auto mode cost extra?

The auto mode classifier uses a small number of extra tokens per tool call. As of August 7, 2026, Anthropic no longer charges Claude Code users on the Pro, Max, and Team plans for that classifier overhead.

How do I change my Claude Code permission mode?

Press Shift+Tab in the CLI to cycle modes, or use the permission mode dropdown in the desktop app and VS Code extension. Administrators can pin an organisation-wide default with defaultMode in managed settings, or switch auto mode off entirely with disableAutoMode.

Is auto mode safer than approving every prompt yourself?

In Anthropic's own controlled study with 1,053 paid testers, a planted dangerous command was caught by human reviewers 13.6% of the time and blocked by auto mode 89% of the time. Anthropic still recommends reviewing Claude's work yourself for high-stakes changes to production infrastructure, because auto mode relies on a classifier and does not eliminate risk.

Lesson 06 · Claude Mastery Pro+ ~16 min Interactive · for developers

Claude Code: the CLI for real work.

Claude Code is Anthropic's command-line tool that gives Claude direct access to your codebase, terminal, and file system. It's the most capable AI coding tool there is — and the riskiest if you let it off the leash. This lesson covers setup, MCP servers, and the review discipline that makes it safe: the patterns that ship code instead of breaking things.

Step 1 of 50% complete

The mental model

Not an autocomplete — a junior who runs commands

Claude Code can read your code, write code, run tests, search the web, and make commits. The difference from Copilot or Cursor is autonomy: it does multi-step work and reports back. Which means the way you use it well looks less like coding and more like managing.

Predict first

What's the right way to think about Claude Code?

Setup & what it does

Install once, then delegate

npm install -g @anthropic-ai/claude-code → run claude in any repo

① Start with low-stakes tasks

Docs, tests, small refactors, codebase Q&A. Ask for a diff before it commits.

"Add JSDoc to every public function in src/utils, run the tests, show me the diff before committing."

② Connect your tools via MCP

GitHub, Linear, Postgres — so it can read issues, query schemas, and see PR context.

Use read-only credentials for any production system — never give it write access to prod.

③ Multi-step feature work

Hand it a brief; it plans, codes, tests, and preps the PR — with you reviewing at checkpoints.

"Plan first, show me, then execute step by step. Ask before destructive operations."

④ Know which permission mode you are in

Claude Code has four: default (reads run free, every edit and command asks), acceptEdits (file edits auto-approved), auto (a classifier decides per tool call), and bypassPermissions (no checks at all — isolated environments only).

Shift+Tab cycles modes in the CLI; use the dropdown on desktop and in VS Code. Admins pin one with defaultMode, or switch auto off with disableAutoMode, in managed settings.

Changed 14 August 2026 — auto mode is the new default. Anthropic is making auto the default permission mode for new Claude Code sessions on the Pro, Max and Team plans. Rather than prompting you before each file write and shell command, a separate classifier screens every tool call and blocks the ones that are irreversible, destructive, or aimed outside your environment; after three blocks in a row, or twenty in a session, Claude Code drops back to manual approvals. If you had already pinned your own default, nothing changes. Anthropic stopped charging Pro, Max and Team users for the classifier’s token overhead on 7 August. Enterprise, the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry stay opt-in for now.

Why the change: Anthropic’s own data showed users approve 97% of permission prompts, and in a controlled study of 1,053 paid testers a planted dangerous command was caught by humans 13.6% of the time versus 89% by auto mode. Worth knowing before you assume the prompts were protecting you — and worth knowing the classifier still misses things, so the habits below do not retire.

Two rules that prevent most disasters: require a plan first on anything multi-step, and never let it commit-and-push in one shot — review the diff yourself.

Do it · review the plan

You're the senior. Approve or flag each step.

You gave Claude Code a brief and asked it to plan first. Here's its plan. Approve the steps that fit the brief; flag the ones that don't or that need a human first.

claude > here's my plan

Brief: Add PATCH /users/:id/delete (sets deleted_at, returns 204), update the User model, add a migration, write unit tests, update the API docs. Don't change existing endpoints. Ask before destructive operations.

The discipline

The senior reads the diff

Claude Code finishes the feature, tests pass, and it offers to commit and push to main in one step to save you time.

Auto mode does not change this answer. A commit and push to a branch you own is exactly the sort of routine action the classifier lets through — it screens for destructive and out-of-environment actions, not for whether the code is any good. Anthropic still recommends reviewing Claude’s work yourself on high-stakes changes.

The call

What's the right move?

⌨️

Lesson complete

You can delegate to Claude Code — and stay the reviewer

What you can do now

Install Claude Code and run your first low-stakes task
Connect your tools via MCP — with read-only creds on anything production
Require a plan first on any multi-step work
Review the plan and the diff like a senior reviewing a junior
Never commit-and-push in one shot; open a PR you can read

Your move: ship one thing with Claude Code

Pick a small feature or chore in your current project. Run it end-to-end — brief, plan, code, test, PR — and review carefully. Notice which parts saved time and which needed your full attention. That calibration is the real skill.

Try a prompt in the playground →

Pro

Up next in Claude Mastery

Custom instructions: teach Claude your style once

How to write custom instructions that genuinely change the output — the four sections every good instruction set needs, and the anti-patterns that make Claude worse.

Continue to Custom instructions → Back to all lessons