Claude Code Permission Hook: Uninterrupted Coding Without the Risk

Problem: Every time Claude wants to read a file or run a command, you're clicking "approve." Twenty clicks later, you've lost your flow state and forgotten what you were building.

Quick Win: Install the Permission Hook and never click approve again:

npm install -g @abdo-el-mobayad/claude-code-fast-permission-hook
cf-approve install
cf-approve config

Three commands. Now Claude runs uninterrupted while dangerous operations get blocked automatically. No --dangerously-skip-permissions required.

The Permission Dilemma

You have two bad options with vanilla Claude Code:

Option 1: Click approve constantly. Safe, but flow-destroying. Complex features mean 50+ permission prompts. You lose context. You lose momentum. You lose the magic of AI-assisted coding.

Option 2: Use --dangerously-skip-permissions. Fast, but terrifying. One hallucinated rm -rf / and your system is gone. Fine for throwaway projects. Unacceptable for real work.

The Permission Hook gives you a third option: intelligent delegation. Claude runs without interruption. Dangerous commands get blocked automatically. Edge cases go to a fast LLM for context-aware decisions.

How It Works: Three-Tier Decision System

When Claude requests permission, the hook evaluates it instantly:

Tier 1 - Fast Approve (No AI Needed)

Safe operations pass through immediately:

• Read, Glob, Grep, WebFetch, WebSearch
• Write, Edit, MultiEdit, NotebookEdit
• TodoWrite, Task, all MCP tools

No latency. No cost. Claude keeps working.

Tier 2 - Fast Deny (No AI Needed)

Dangerous operations get blocked instantly:

# These never execute, period
rm -rf /                      # System destruction
git push --force origin main  # Protected branch overwrite
mkfs /dev/sda                 # Disk formatting
:(){ :|:& };:                  # Fork bomb

No AI evaluation needed. Hard-coded rules protect you from catastrophic mistakes.

Tier 3 - LLM Analysis (Cached)

Ambiguous operations get sent to a fast, cheap LLM (GPT-4o-mini via OpenRouter) for context-aware evaluation:

{
  "tool": "Bash",
  "command": "docker system prune -af",
  "working_directory": "/home/user/project",
  "recent_context": "User asked to clean up Docker resources"
}

The LLM sees what you're trying to accomplish and makes an intelligent decision. Decisions are cached - repeat the same command and it's instant.

Configuration

The hook stores settings at ~/.claude-code-fast-permission-hook/config.json:

{
  "llm": {
    "provider": "openai",
    "model": "openai/gpt-4o-mini",
    "apiKey": "sk-or-v1-your-key",
    "baseUrl": "https://openrouter.ai/api/v1"
  },
  "cache": {
    "enabled": true,
    "ttlHours": 168
  }
}

OpenRouter is recommended for best latency. Get your key at openrouter.ai. Cost: roughly $1 per 5,000+ LLM decisions. In practice, most operations hit Tier 1 or 2, so a dollar lasts months.

Installation Levels

Device Level (recommended): Configure once in ~/.claude/settings.json, applies everywhere. Set it and forget it.

Project Level: Configure in .claude/settings.local.json for project-specific rules.

The installer adds this to your settings:

{
  "hooks": {
    "PermissionRequest": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "cf-approve permission"
          }
        ]
      }
    ]
  }
}

When Things Go Wrong

Error: "Permission denied" on all operations

Fix: Your API key is missing or invalid:

cf-approve config

Re-enter your OpenRouter key.

Error: "Hook not triggering"

Fix: Verify installation:

cf-approve doctor
cf-approve status

Behavior seems inconsistent

Fix: Clear the decision cache:

cf-approve clear-cache

The Two Foundational Hooks

The Permission Hook is one half of ClaudeFast's hook philosophy. The other half is the Skill Activation Hook, which ensures Claude loads the right skills at the right time.

Together, they create friction-free development: you type naturally, Claude works efficiently, and the framework handles orchestration invisibly.

You Can Now Code Without Interruption

• You just installed context-aware permission automation
• Try this next: Learn about Claude Code configuration for more optimization
• Go deeper: Explore skills for specialized agent workflows

No more permission fatigue. No more dangerous flags. Just Claude doing what Claude does best - building your software while you focus on the big picture.