aludra runs on each Mac or Linux machine whose projects I want to reach. It runs Claude Code sessions in a project on that machine, using the Claude Code that’s already installed and logged in there, and keeps them in a Pi Durable SQLite store. A session keeps going when every client is closed and picks up where it was after the host restarts.
I reach it from amyrlin, a native macOS app that shows sessions from several machines at once, from the aludra command on the machine itself, or through its HTTP API. While a run is going I can steer it, which takes effect at the next tool call, or queue a follow-up for when it answers.
The host listens only on localhost. tailscale serve is the way in, and only the Tailscale logins I list get through. Inside a project the agent works without asking, but git push and SSH to hosts I’ve listed wait for my approval, and credential files are blocked. Those checks read shell commands as text, so they catch mistakes, not an agent that’s trying to get around them.