Approve your AI coding agent's risky commands from your phone — so it keeps working while you're away.
When Claude Code / Codex / Gemini CLI / Cursor needs permission to run something risky
(force-push, rm -rf, a deploy, sending email), Push Bridge sends the request to your phone.
You see the exact command, tap Approve / Deny, and your agent continues.
The whole promise of agentic coding is "it works while you do other things" — until it stops and waits for you to approve one command. Push Bridge closes that gap.
- Claude Code — a
PreToolUsehook forwards the prompt to a relay and long-polls for your decision. - Any MCP agent (Codex, Gemini CLI, Cursor, …) — a
request_approvaltool that blocks until you respond. - Phone side — a PWA using Web Push. You get the notification; you decide.
A remote approve is also a remote kill switch — so you can run agents more autonomously without losing control. Human-in-the-loop, but the human can be anywhere.
curl -fsSL https://pwa-artvibes-projects.vercel.app/install.sh | bashThen add the hook to ~/.claude/settings.json (the installer prints the exact snippet). See plugin/README.md.
See mcp/README.md — exposes request_approval and notify tools to any MCP client.
- Free / OSS (this repo, MIT): the plugin + MCP server. Bring your own webhook (e.g. ntfy.sh) — get pinged when your agent needs you.
- Pro ($9/mo): hosted relay + the phone app with Approve/Deny. → https://pwa-artvibes-projects.vercel.app
iOS web push doesn't render notification action buttons yet, so on iOS you tap the notification and Approve/Deny inside the app (one extra tap). Android & desktop get the buttons directly on the notification.
MIT — see LICENSE.