Claude Code — Hooks

Run AgentGuards checks as Claude Code hooks to enforce guardrails at the tool-execution boundary. Hooks fire before Claude executes shell commands and before Claude responds to user prompts.

Hook types

Hook	When it fires	Enforcement
PreToolUse (Bash)	Before any shell command executes	Hard block — command never runs
UserPromptSubmit	After user submits a prompt	Soft block — Claude Code won't respond

Note: UserPromptSubmit hooks cannot intercept the LLM call to Anthropic — they only prevent Claude from showing a response. For message-level interception, use the proxy integration.

Setup

1. Install the hook script

bash

cp scripts/agentguards_hook.py ~/.claude/agentguards_hook.py

On Windows: copy scripts\agentguards_hook.py %USERPROFILE%\.claude\agentguards_hook.py

2. Configure hooks in ~/.claude/settings.json

~/.claude/settings.json

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "AGENTGUARDS_URL=https://prod.agentguards.co AGENTGUARDS_API_KEY=ag_YOUR_TOKEN python3 ~/.claude/agentguards_hook.py UserPromptSubmit"
          }
        ]
      }
    ],
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "AGENTGUARDS_URL=https://prod.agentguards.co AGENTGUARDS_API_KEY=ag_YOUR_TOKEN python3 ~/.claude/agentguards_hook.py PreToolUse"
          }
        ]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Bash|WebFetch|WebSearch|Write|Edit|MultiEdit",
        "hooks": [
          {
            "type": "command",
            "command": "AGENTGUARDS_URL=https://prod.agentguards.co AGENTGUARDS_API_KEY=ag_YOUR_TOKEN python3 ~/.claude/agentguards_hook.py PostToolUse"
          }
        ]
      }
    ]
  }
}

Recommended: set the env vars in your shell profile so the API key is not visible in process listings.

~/.zshrc

# ~/.zshrc or ~/.bashrc
export AGENTGUARDS_URL=https://prod.agentguards.co
export AGENTGUARDS_API_KEY=ag_YOUR_TOKEN

Then the settings become:

~/.claude/settings.json

{
  "hooks": {
    "UserPromptSubmit": [
      { "hooks": [{ "type": "command", "command": "python3 ~/.claude/agentguards_hook.py UserPromptSubmit" }] }
    ],
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [{ "type": "command", "command": "python3 ~/.claude/agentguards_hook.py PreToolUse" }]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Bash|WebFetch|WebSearch|Write|Edit|MultiEdit",
        "hooks": [{ "type": "command", "command": "python3 ~/.claude/agentguards_hook.py PostToolUse" }]
      }
    ]
  }
}

3. Verify

Run a Claude Code session and try a blocked command. The hook calls POST /v1/actions/authorize and blocks the command before it executes.

test

User: Run rm -rf /

Understanding hook behaviour

Want a plain-language explanation of what each hook does, which commands are approved or blocked, and what the decisions mean?

How hooks work →

Combining with other integrations

Proxy

Intercepts prompts before they reach Anthropic — message-level enforcement.

MCP

Exposes guardrail tools Claude can call cooperatively from its reasoning.

Proxy + Hooks together cover both conversation-level and action-level threats — the recommended setup for API key users. For Claude Pro / Max subscribers (no proxy available), hooks are the primary enforcement mechanism for tool execution.

Fail-open behaviour

The hook script is designed to fail open: if the AgentGuards service is unreachable (network issue, timeout), the hook exits 0 and allows the action. This prevents the guardrail service from becoming a hard dependency that blocks your work.

To change this to fail-closed, edit the _allow() calls in the except blocks of agentguards_hook.py.