Permission Modes

Why permissions matter

Claude Code has real access to your filesystem and shell. It can read files, write files, run commands, and make network requests. Permissions are how you control the blast radius.

The permission system has two layers:

1. Modes - preset permission levels you can switch between
2. Settings - granular per-tool allow/ask/deny lists

The five permission modes

default

The starting mode. Claude reads freely but asks permission before writing files or running commands. You see a prompt like:

Claude wants to edit src/auth.ts
[Allow] [Deny] [Always allow]

Good for: learning Claude Code, working in unfamiliar codebases, pair programming.

acceptEdits

File edits are auto-accepted. Shell commands still require approval. This is the sweet spot for most development work - you trust Claude's code changes but want to approve any rm, git push, or npm install.

plan

Read-only. Claude can read your codebase and propose a plan, but cannot modify anything. Use this when you want Claude to analyze and suggest without touching anything.

claude --permission-mode plan

> How would you refactor the auth module?

Claude will read files, analyze the code, and propose a detailed plan
with specific file changes - but won't execute any of them.

dontAsk

Everything is auto-approved. Claude reads, writes, and runs commands without asking. Use this when you trust the task and want Claude to work autonomously.

bypassPermissions

Like dontAsk but also bypasses system-level restrictions. Only available via CLI flag:

claude --dangerously-skip-permissions

Switching modes

In interactive mode: press Shift+Tab to cycle through modes. The current mode shows in the prompt indicator.

Via CLI flag:

claude --permission-mode plan
claude --permission-mode acceptEdits
claude --dangerously-skip-permissions

The "Always allow" shortcut: when Claude asks for permission in default mode, choosing "Always allow" adds that specific tool to your allow list for the session.

Settings-based permissions

For finer control, configure permissions in .claude/settings.json (shared with team, committed to git) or .claude/settings.local.json (personal, gitignored):

{
  "permissions": {
    "allow": [
      "Read",
      "Glob",
      "Grep",
      "Bash(npm test)",
      "Bash(npm run lint)"
    ],
    "deny": [
      "Bash(rm -rf *)",
      "Bash(git push --force)"
    ]
  }
}

The three arrays:

• allow - always permitted, no prompt
• ask - prompt before each use (default for most tools)
• deny - always blocked, even if the mode would allow it

You can be specific with shell commands. Bash(npm test) allows only that exact command. Bash(npm run *) allows any npm run script.

Recommended setup for teams

Share safety rails in .claude/settings.json (committed to git):

{
  "permissions": {
    "allow": [
      "Read",
      "Glob",
      "Grep",
      "Bash(npm test)",
      "Bash(npm run lint)",
      "Bash(npm run build)"
    ],
    "deny": [
      "Bash(rm -rf *)",
      "Bash(git push --force)",
      "Bash(curl * | sh)"
    ]
  }
}

Individual developers can add personal overrides in .claude/settings.local.json (gitignored).

{
  "permissions": {
    "allow": [
      "Read",
      "Glob",
      "Grep",
      "Bash(npm test)",
      "Bash(npm run lint)",
      "Bash(npm run build)",
      "Bash(npx tsc --noEmit)"
    ],
    "deny": [
      "Bash(git push --force*)",
      "Bash(rm -rf *)"
    ]
  }
}