Stop AI agents from taking actions you didn't approve.
AgentMFA adds a human approval step to your AI agent. Before it deletes files, deploys code, sends emails, or does anything sensitive — it pauses and asks you first. You approve or reject from your phone with FaceID or fingerprint.
brew install agentmfa/tap/agentmfaOr download directly from github.com/agentmfa/agentmfa/releases.
agentmfa auth loginThis opens a browser window to sign in with Google. Your session is stored securely in the system keychain — no API keys or environment variables needed.
claude plugin marketplace add https://github.com/agentmfa/agentmfa-integrations
claude plugin install agentmfa@agentmfaagentmfa agent registerApprove the registration on your phone. This links the Claude instance on this machine to your account and generates a cryptographic keypair for signed requests.
Download the AgentMFA app on your phone and sign in with the same account. This is where you'll receive approval requests.
That's it — Claude will now ask for your approval before taking any sensitive action.
brew install agentmfa/tap/agentmfa
agentmfa auth login
agentmfa agent registerAdd this to your ~/.openclaw/openclaw.json:
{
"mcp": {
"servers": {
"agentmfa": {
"command": "agentmfa",
"args": ["serve"]
}
}
}
}Via ClawHub (recommended):
openclaw skills install agentmfaVia curl — global, available to all your OpenClaw agents:
curl -fsSL https://raw.githubusercontent.com/agentmfa/agentmfa-integrations/main/install-skill.sh | shVia curl — per workspace, available only in a specific project:
curl -fsSL https://raw.githubusercontent.com/agentmfa/agentmfa-integrations/main/install-skill.sh | sh -s -- /path/to/your/workspace/skillsDownload the AgentMFA app on your phone and sign in with the same account.
AgentMFA works with any MCP-compatible agent (Cursor, Windsurf, etc.). Install the CLI, log in, register the agent, then add this to your MCP config:
{
"mcpServers": {
"agentmfa": {
"command": "agentmfa",
"args": ["serve"]
}
}
}- You ask Claude to do something that could have consequences (deploy to production, delete data, send an email, etc.)
- Claude pauses and sends a notification to your phone
- You see what it's about to do and tap Approve or Reject
- Claude proceeds — or stops — based on your decision
