My main agent is Claude Code but I've been dabbling with Copilot CLI, Cursor CLI, and Codex CLI too. The biggest productivity killer across all of them is the same: agents blocking on permission prompts while I'm away from my desk.
So I made Greenlight: an iOS app that forwards agent permission requests to your phone as push notifications. And it worked! And then Anthropic released "Remote Control". That was a tough day. But I realized something: there's no agent-agnostic remote control. So I've extended Greenlight to work with Copilot, Cursor and Codex too.
The companion CLI (`greenlight connect`) wraps your agent session. The agent runs in full auto mode while Greenlight intercepts actions before they execute. This inverts control: instead of the agent deciding what permissions to ask you for, you decide what actions the agent is permitted to take. Did you know that Claude Code runs `ssh ... [email protected]` every time you start a session?
Other stuff I like about it: compound shell commands are split into segments and color-coded by risk level so it's easy to judge the risk of a command. Over time your rules are tuned to the project and you only receive requests for novel or destructive commands. My favorite feature is "pull the plug": you can sigkill any agent that is rabbitholing and ignoring you.
The server mostly a router - it never inspects or saves your transcript. Would love feedback, especially from anyone juggling multiple agents.