Setup Guide

Five minutes. Let's go.

Install the App

Download ClawGate from the App Store and create your account.

Download for iOS

When you open the app for the first time:

Create your account with email
You'll be prompted to register a passkey — just scan your face. That's it.
Your passkey syncs via iCloud Keychain across all your Apple devices

Get Your API Key

In the app, go to Settings → API Keys → Create New.

Give it a name (e.g. "My Agent") and tap Create. Your key will look like:

ag_xDWZ7V_OamQEDz-uBk5t4LeM6bZxBHXi

Copy it now — you won't be able to see it again. (You can always generate a new one.)

Install the Plugin

Install the ClawGate plugin for OpenClaw:

openclaw plugins install clawgate

Or install manually: npm install clawgate in your OpenClaw extensions directory (~/.openclaw/extensions/clawgate/).

Configure OpenClaw

Add ClawGate to your OpenClaw config file (~/.openclaw/openclaw.json):

{
  "plugins": {
    "entries": {
      "clawgate": {
        "enabled": true,
        "config": {
          "apiKey": "ag_YOUR_KEY_HERE",
          "agentId": "my-agent"
        }
      }
    }
  }
}

The server URL defaults to https://api.clawgate.ai — no need to set it unless you're self-hosting.

Restart OpenClaw to load the plugin:

openclaw gateway restart

Test It

Run the built-in test command:

openclaw clawgate test

This sends a test approval request to your phone. Approve it with Face ID to confirm everything works.

Then try a real interception — ask your agent to delete a test file:

# Create a test file, then ask your agent:
# "Delete the file test.txt"
# You'll get a push notification — approve or deny with Face ID

🎉

You're all set

ClawGate automatically intercepts sensitive tool calls at the infrastructure level. Your agent can't bypass it — approve or deny with Face ID in under 2 seconds.

Configuring Sensitive Patterns

ClawGate comes with sensible defaults that catch dangerous operations: file deletion (rm, Remove-Item, del), disk operations, npm publish, git push --force, database DROP/TRUNCATE, and more.

To customize patterns, edit the patterns array in your OpenClaw config:

{
  "plugins": {
    "entries": {
      "clawgate": {
        "enabled": true,
        "config": {
          "apiKey": "ag_YOUR_KEY_HERE",
          "agentId": "my-agent",
          "patterns": [
            { "match": "rm\\s", "description": "Delete files (Unix)", "tools": ["exec"] },
            { "match": "Remove-Item", "description": "Delete files (PowerShell)", "tools": ["exec"] },
            { "match": "send.*email", "description": "Send email" },
            { "match": "stripe.*charge", "description": "Create Stripe charge" },
            { "match": "npm publish", "description": "Publish to npm", "tools": ["exec"] },
            { "match": "git push.*--force", "description": "Force push to git", "tools": ["exec"] },
            { "match": "drop.*table", "description": "Drop database table", "tools": ["exec"] }
          ]
        }
      }
    }
  }
}

Patterns use regular expressions (case-insensitive). The tools array limits a pattern to specific tool names (e.g. only exec). If omitted, the pattern matches all tools.

You can also require approval for all calls to a specific tool:

"alwaysRequireApproval": ["message", "browser"]

View your active patterns anytime: openclaw clawgate patterns

For AI Agents

Good news: you don't need to do anything special. ClawGate works automatically.

🔒 How It Works (v2 — Automatic Enforcement)

ClawGate v2 uses OpenClaw's before_tool_call hook to intercept tool calls at the infrastructure level. This means:

No agent code changes needed — the plugin intercepts automatically
Agents cannot bypass it — enforcement happens before the tool executes
Pattern matching determines which calls need approval
Fail-closed by default — if the server is unreachable or the request times out, the action is blocked

When a sensitive tool call is detected, the agent's execution pauses while ClawGate sends a push notification to the user's phone. The user approves or denies with Face ID. If denied (or timed out), the tool call is blocked and the agent receives a denial message.

💡 API Quick Reference

If you need to integrate ClawGate outside of OpenClaw, here's the API:

POST https://api.clawgate.ai/api/v1/approval/request
Authorization: Bearer ag_USER_KEY_HERE
Content-Type: application/json

{
  "agentId": "my-agent",
  "action": "send-email",
  "description": "Send quarterly report to investors@company.com",
  "timeoutSeconds": 120
}

// Poll for result:
GET https://api.clawgate.ai/api/v1/approval/status/{requestId}

// Response: { "status": "APPROVED" | "DENIED" | "PENDING" | "EXPIRED" }

Need help? Check the full docs or open an issue.

Full Documentation GitHub Issues