Hey hunters! 👋 I'm Serge, the maker of Agents At Work. The problem: I kick off Claude Code on a long-running task, then walk away from my Mac. Every time, the same thing happened. I'd come back an hour later to find the agent stuck on a permission prompt it asked ten minutes in, just sitting there waiting for a yes/no I never saw. Or it had crashed silently. The "let it run autonomously" promise of these CLIs falls apart the moment you're not at the keyboard to babysit them. I started building this for myself before Claude shipped its own remote control and mobile features. I figured Anthropic would eventually build something like it, and they did. But once I had my own bridge running, it was already doing more: full conversation context, detailed permission prompts, and it wasn't locked to one agent. I switch between Claude Code, Codex, Gemini, Grok, Cursor, and Scoot depending on the task, and I didn't want a separate babysitting workflow for each one. That's when I decided to generalize it and make it public instead of keeping it to myself. A few things I looked at before building this: Claude.ai mobile lets you dispatch tasks to Claude while your Mac is awake, which is great for simple jobs. But it only controls Claude, gives limited real-time visibility, has simpler permission prompts, and no conversation feed. No Codex, Gemini, Grok, Cursor, or local models. Generic push tools (Pushover, ntfy, and the like) just send a notification. No context, no way to reply or approve from the alert itself. OpenClaw does remote control over Telegram for quick commands, but setup is involved and there's no dedicated UI showing full session context. So I built the bridge I wanted. Hooks on the agent side write events straight to Firestore, a small daemon keeps a heartbeat and listens for commands, and the phone app shows a real conversation feed instead of a beep. When the agent hits a tool call that needs approval, you get a push notification with the actual question and tappable options. Answer from your phone, and the daemon sends the keystroke back into the terminal session so the agent continues. You can also drop a note mid-session, like "skip the tests for now," without opening a laptop. End-to-end encryption (AES-256-GCM) went in once I realized I was piping real prompts and file paths through a third-party backend, and I wasn't comfortable with that myself. If I wouldn't use it that way, I wasn't going to ship it to anyone else. It supports Claude Code, Codex CLI, Gemini CLI, Grok, Cursor, and Scoot (my own open-source agent that runs free local models with Ollama and cloud models behind one interface). Native iOS and Android apps, plus a headless option that works on both macOS and Linux, so a home server or cloud box is a fine place to leave an agent working overnight. Auth is QR-only on mobile: scan and go, no passwords. One recent addition I'm proud of: scheduled prompts. Queue a prompt now and have the agent pick it up at a set time, say 2 AM when your token limit resets, so it works overnight while you sleep instead of sitting idle against a spent quota. You set it from the phone, the Mac, or the Linux CLI, and you can edit or cancel it any time before it fires. How do you handle unattended agent sessions today? Happy to dig into the architecture, the encryption approach, or why I built one bridge for six agents instead of one.

Agents At Work

Monitor and control AI agents from your phone

Monitor and control AI agents from your phone

Agents At Work: You can go!

Previous Agents At Work Launches

Previous Agents At Work Launches