tuchg/lucarne

Lucarne is a small background service for your computer that pipes the chatter of local AI coding agents out to your phone, so you can leave the laptop and still get pinged when one of them needs an answer. The agents it supports today are Claude, Codex, Gemini, Copilot, and Pi. The pitch on the front of the README is direct: stop babysitting your local AI agents. The selling point is that it does not change how those agents are installed or what they have access to. There are no hooks, no skills, no MCP plugins, no edits to the project you are working on. You install Lucarne with Homebrew, run an init command, and either point it at a Telegram bot or scan a WeChat QR code. From that moment, when one of your agents asks for permission to run a command, asks a clarifying question, or fails, the message shows up as a chat notification you can tap to approve or deny. The two channels feel slightly different in use. In WeChat, you quote one of the notifications and reply to it, and Lucarne uses the quoted message to figure out which agent session you are talking to, so the reply lands back in the right context. Lucarne tries to match by the original message id first, then by a hash of the quoted text. In Telegram, the entry point is a /panel command which gives you a mobile console of all your workspaces and sessions, mapped to Telegram forum topics, with buttons for approve, deny, status, interrupt, and fork. Under the hood the code is written in Rust and structured as separate crates. A channel adapter layer talks to Telegram or WeChat. The core holds the runtime bus, the control plane, history, and the daemon. A separate piece called agent-sessions watches and parses the on disk session files of each supported agent CLI. Capability varies by agent: Claude and Codex get the full set including sub-agents and forking sessions, while Copilot is more limited. The project is MIT licensed, supports macOS today, and lists Linux and Windows on the roadmap.