January 12, 20269 MIN READ

Claude Code Permissions: Deny, Allow & Ask Modes Explained

By Learnia Team

Claude Code Permissions: Deny, Allow & Ask Modes Explained

This article is written in English. Our training modules are available in French.

Claude Code can read files, execute commands, modify your codebase, and interact with external services. That power requires control. The permission system gives you granular authority over what Claude can do—from fully supervised sessions to autonomous automation.

Why Permissions Matter

Unlike chat-based assistants, Claude Code is an agentic tool. It doesn't just suggest—it acts. Without guardrails:

→A misunderstood request could delete files
→Tests might run against production
→Credentials could be logged or exposed
→Destructive commands could execute silently

The permission system prevents these scenarios while preserving Claude's ability to help.

The Three Permission Modes

Claude Code operates in one of three modes for each tool or command:

Mode	Behavior	Use Case
Ask	Prompts before each action	Interactive development
Allow	Executes without prompting	Trusted operations
Deny	Blocks completely	Dangerous commands

Ask Mode (Default)

Ask mode is the safest interactive option. Claude shows you what it wants to do and waits for approval:

Claude wants to run: npm install lodash

[y] Accept  [n] Reject  [e] Edit  [Esc] Cancel

You can:

→Accept (y): Run the command
→Reject (n): Skip this action
→Edit (e): Modify the command before running
→Cancel (Esc): Cancel entirely

Allow Mode

Allow mode permits specific operations without prompts:

# Allow all npm commands
claude config add permissions.allow "Bash(npm:*)"

# Allow git operations  
claude config add permissions.allow "Bash(git:*)"

# Allow file reads
claude config add permissions.allow "Read(*)"

Claude runs allowed commands immediately—no confirmation needed.

Deny Mode

Deny mode blocks operations completely:

# Block production database access
claude config add permissions.deny "Bash(psql:*prod*)"

# Block rm with -rf flag
claude config add permissions.deny "Bash(rm:-rf*)"

# Block curl to external URLs
claude config add permissions.deny "Bash(curl:*)"

Even if Claude tries to run a denied command, it fails immediately.

Permission Precedence

When multiple rules apply, Claude evaluates them in this order:

→Deny — Checked first. If a deny rule matches, action blocked.
→Allow — Checked second. If an allow rule matches, action permitted.
→Ask — Default. If no rule matches, prompt the user.

Example:

# Allow all git commands
claude config add permissions.allow "Bash(git:*)"

# But deny force push
claude config add permissions.deny "Bash(git:push --force*)"
claude config add permissions.deny "Bash(git:push -f*)"

Result: Most git commands run freely, but force pushes are blocked.

Tool Types & Permission Patterns

Built-in Tools

Tool	Description	Pattern Format
`Read`	Read file contents	`Read(path/pattern)`
`Write`	Write/modify files	`Write(path/pattern)`
`Edit`	Edit files	`Edit(path/pattern)`
`Bash`	Execute shell commands	`Bash(command:args)`
`Browser`	Browse URLs	`Browser(url-pattern)`
`MCP`	MCP server tools	`mcp__servername__toolname(*)`

Pattern Syntax

Patterns use glob-style matching:

Pattern	Matches
`*`	Any single segment
`**`	Any path depth
`?`	Single character
`{a,b}`	Either a or b

Examples:

# Allow reading any file
"Read(*)"

# Allow reading only src folder
"Read(src/**)"

# Allow npm and yarn
"Bash({npm,yarn}:*)"

# Allow git except specific commands
"Bash(git:*)" + deny "Bash(git:push --force*)"

Configuring Permissions

Interactive Configuration

> /permissions

Opens an interactive menu:

Current Permission Mode: ask

Allowed:
  - Bash(npm:*)
  - Bash(git:*)
  - Read(*)

Denied:
  - Bash(rm:-rf*)
  - Bash(sudo:*)

[a] Add rule  [r] Remove rule  [m] Change mode  [Esc] Exit

Command-Line Configuration

# Add allow rules
claude config add permissions.allow "Bash(npm:*)"
claude config add permissions.allow "Bash(git:*)"

# Add deny rules
claude config add permissions.deny "Bash(rm:-rf*)"

# View current permissions
claude config show permissions

# Remove a rule
claude config remove permissions.allow "Bash(npm:*)"

Settings File

Permissions live in

~/.claude/settings.json

{
  "permissions": {
    "mode": "ask",
    "allow": [
      "Bash(npm:*)",
      "Bash(git:*)",
      "Read(*)",
      "Edit(src/**)"
    ],
    "deny": [
      "Bash(rm:-rf*)",
      "Bash(sudo:*)",
      "Bash(curl:*)"
    ]
  }
}

Project-Level Permissions

Add permissions to your project's

.claude/settings.json

{
  "permissions": {
    "allow": [
      "Bash(npm:*)",
      "Bash(pytest:*)"
    ],
    "deny": [
      "Bash(npm:publish*)"
    ]
  }
}

Project permissions merge with user permissions. Deny rules from either source take precedence.

Trust Levels

For quick configuration, use trust levels:

Default (Ask Everything)

claude config set permissions.mode "ask"

All operations require confirmation.

Auto-Accept Edits

claude config set permissions.mode "auto-edit"

File edits are automatically accepted; other operations still ask.

Full Auto

claude config set permissions.mode "auto"

All safe operations proceed automatically. Only potentially dangerous commands prompt.

Sandbox Mode

claude --sandbox

Runs in isolated environment:

→Filesystem restricted to project
→Network limited to localhost
→No access to system commands

See Claude Code Best Practices: Security, Performance & Teams for sandbox details.

Session-Based Permissions

Grant temporary permissions for specific sessions:

Session Trust

When Claude requests an action, you can grant session-wide permission:

Claude wants to run: npm test

[y] Yes  [Y] Yes, allow all npm commands this session
[n] No   [Esc] Cancel

Pressing

(capital) adds a temporary allow rule for the session.

Dangerously Skip Permissions

For fully autonomous execution (CI/CD, automation):

claude --dangerously-skip-permissions

Warning: This bypasses all permission checks. Only use in:

→Sandboxed environments
→CI/CD pipelines
→Fully controlled contexts

See Claude Code GitHub Actions: AI-Powered CI/CD Automation.

Common Permission Configurations

Web Developer

{
  "permissions": {
    "allow": [
      "Read(*)",
      "Edit(src/**)",
      "Edit(public/**)",
      "Bash(npm:*)",
      "Bash(yarn:*)",
      "Bash(pnpm:*)",
      "Bash(git:*)",
      "Bash(eslint:*)",
      "Bash(prettier:*)"
    ],
    "deny": [
      "Bash(rm:-rf*)",
      "Bash(npm:publish*)",
      "Edit(.env*)"
    ]
  }
}

Python Data Scientist

{
  "permissions": {
    "allow": [
      "Read(*)",
      "Edit(src/**)",
      "Edit(notebooks/**)",
      "Bash(python:*)",
      "Bash(pip:*)",
      "Bash(pytest:*)",
      "Bash(jupyter:*)"
    ],
    "deny": [
      "Bash(rm:-rf*)",
      "Edit(.env*)",
      "Bash(pip:install --user*)"
    ]
  }
}

DevOps Engineer

{
  "permissions": {
    "allow": [
      "Read(*)",
      "Edit(terraform/**)",
      "Edit(ansible/**)",
      "Bash(terraform:*)",
      "Bash(ansible:*)",
      "Bash(kubectl:get*)",
      "Bash(kubectl:describe*)"
    ],
    "deny": [
      "Bash(kubectl:delete*)",
      "Bash(terraform:destroy*)",
      "Bash(rm:-rf*)",
      "Edit(**/secrets/**)"
    ]
  }
}

CI/CD Automation

{
  "permissions": {
    "allow": [
      "Read(*)",
      "Edit(*)",
      "Bash(npm:*)",
      "Bash(git:*)",
      "Bash(gh:*)"
    ],
    "deny": [
      "Bash(rm:-rf /)",
      "Bash(sudo:*)"
    ]
  }
}

Used with

--dangerously-skip-permissions

in controlled CI environments.

Security Best Practices

1. Start Restrictive

Begin with default ask mode. Add allow rules only for operations you trust completely:

# Start here
claude config set permissions.mode "ask"

# Add specific allows as needed
claude config add permissions.allow "Bash(npm:install)"
claude config add permissions.allow "Bash(npm:test)"

2. Always Deny Destructive Commands

claude config add permissions.deny "Bash(rm:-rf*)"
claude config add permissions.deny "Bash(sudo:*)"
claude config add permissions.deny "Bash(chmod:777*)"
claude config add permissions.deny "Bash(dd:*)"

3. Protect Sensitive Files

# Deny editing environment files
claude config add permissions.deny "Edit(.env*)"
claude config add permissions.deny "Edit(**/.env*)"

# Deny reading credentials
claude config add permissions.deny "Read(**/credentials*)"
claude config add permissions.deny "Read(**/*.pem)"

4. Restrict Production Access

# Block production database commands
claude config add permissions.deny "Bash(psql:*prod*)"
claude config add permissions.deny "Bash(mysql:*production*)"

# Block production deployments
claude config add permissions.deny "Bash(kubectl:*--context=prod*)"
claude config add permissions.deny "Bash(ssh:*prod*)"

5. Use Project-Specific Rules

Team projects should define permissions in

.claude/settings.json

{
  "permissions": {
    "deny": [
      "Bash(npm:publish*)",
      "Bash(git:push --force*)",
      "Edit(.env*)"
    ]
  }
}

Commit this to version control so all team members share the same safety rules.

Hooks for Permission Control

Hooks provide another layer of control. Create custom validation:

// .claude/settings.json
{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "command": "python .claude/hooks/validate-command.py"
      }
    ]
  }
}

The hook script receives command details and can block execution:

# .claude/hooks/validate-command.py
import json
import sys

input_data = json.loads(sys.stdin.read())
command = input_data.get("tool_input", {}).get("command", "")

# Block any command targeting /etc
if "/etc" in command:
    result = {
        "decision": "block",
        "reason": "Commands targeting /etc are not allowed"
    }
else:
    result = {"decision": "proceed"}

print(json.dumps(result))

Learn more: Claude Code Hooks: Automate Your Development Workflow

Debugging Permissions

View Applied Rules

claude config show permissions

Test a Pattern

Check if a specific command would be allowed:

claude permissions check "Bash(npm:publish)"

Output:

Command: Bash(npm:publish)
Result: DENIED

Matching rule:
  deny: Bash(npm:publish*)
  Source: ~/.claude/settings.json

View Permission Logs

Enable verbose logging:

claude --verbose

Shows permission checks in real-time:

[PERMISSION] Checking: Bash(npm:test)
[PERMISSION] Rule matched: allow Bash(npm:*)
[PERMISSION] Decision: ALLOW

Permissions in Automation

GitHub Actions

- name: Run Claude Code
  run: |
    claude --dangerously-skip-permissions -p "Fix failing tests"
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

See Claude Code GitHub Actions: AI-Powered CI/CD Automation.

Headless Mode

claude --print --dangerously-skip-permissions -p "Analyze this codebase"

See Headless & Programmatic Claude Code: SDK & Automation.

MCP Tool Permissions

MCP servers add custom tools. Control them with patterns:

# Allow all tools from filesystem server
claude config add permissions.allow "mcp__filesystem__*"

# Deny destructive filesystem operations
claude config add permissions.deny "mcp__filesystem__delete_file"

# Allow specific GitHub operations
claude config add permissions.allow "mcp__github__get_pull_request"
claude config add permissions.deny "mcp__github__merge_pull_request"

Learn more: Model Context Protocol (MCP) for Claude Code: Complete Guide

Key Takeaways

→
Default to Ask mode: Start supervised and grant permissions incrementally.
→
Deny before Allow: Deny rules always take precedence, providing hard blocks on dangerous operations.
→
Use patterns wisely: Glob patterns give flexibility—be specific to avoid over-permitting.
→
Protect sensitive data: Always deny access to credentials, environment files, and production systems.
→
Commit project permissions: Share safety rules via
```
.claude/settings.json
```
in version control.

Master AI Safety & Ethics

Permissions are just one aspect of responsible AI use. Learn the broader principles in our security module.

In our Module 8 — Safety & Ethics, you'll learn:

→AI safety principles and best practices
→Responsible disclosure and risk management
→Ethical considerations in AI development
→Building trust through transparency

→ Explore Module 8: Safety & Ethics

GO DEEPER

Module 8 — Ethics, Security & Compliance

Navigate AI risks, prompt injection, and responsible usage.

Explorer le Module