Claude Auto Approve — AI skill for Claude Code

claude-auto-approve

A Claude Code [hook](https://docs.anthropic.com/en/docs/claude-code/hooks) that automatically approves safe Bash commands and WebFetch URLs using a cheap/free LLM API, so you only get permission prompts for potentially dangerous operations.

How it works

When Claude Code wants to run a Bash command or fetch a URL, this hook intercepts the permission request and asks a cheap LLM whether the operation could cause harm. Safe commands like `ls`, `git status`, or `cat` and safe URLs like `https://docs.python.org` are auto-approved. Dangerous commands like `rm -rf` or dangerous URLs like `http://192.168.1.1/admin` trigger the normal permission prompt.

Provider failover

The script tries multiple API providers in order:

1. OpenCode — Kimi K2.5, free (~1.5s latency)
2. Moonshot — Kimi K2.5, cheap (~1s latency)

If a provider returns a rate limit error, it enters a 30-minute cooloff period and the next provider is tried.

Fail-safe design

If all providers fail, time out, or return an unrecognizable response, the hook outputs nothing — which tells Claude Code to show the normal permission prompt. You never accidentally approve a dangerous command due to an API error.

Classification prompts

For **Bash commands**, the LLM is asked: "Should a system administrator be warned before running this command?" For **WebFetch URLs**, it checks for SSRF risks (private/internal IPs), data exfiltration patterns, and malicious domains. Both must answer YES or NO.

Requirements

• curl
• jq
• API keys from one or more providers (see below)

Installation

1. Copy the script to your Claude Code hooks directory:

   mkdir -p ~/.claude/hooks
   cp auto-approve-safe.sh ~/.claude/hooks/
   chmod +x ~/.claude/hooks/auto-approve-safe.sh

2. Edit the script and replace the placeholder API keys with your own:

   "moonshot|https://api.moonshot.ai/v1/chat/completions|YOUR_MOONSHOT_API_KEY|kimi-k2.5|0.60|3.00"
   "