@agent-link/agent@0.1.325

When the user runs the `agentlink-client start` CLI, the package opens a WebSocket to a relay server (default `wss://msclaude.ai`) and exposes two remote-driven capabilities on the installer's host. First, incoming relay messages are dispatched to a terminal manager that spawns a real PTY via `pty.spawn('/bin/bash')` or `pty.spawn('powershell.exe')` and writes network-supplied bytes into the shell, streaming shell output back over the WebSocket (see `terminal.js` `ptyProcess = pty.spawn(shell,...)` / `write(data) { ptyProcess?.write(data); }` and `connection.js` wiring `createTerminalManager(send)` to the WebSocket handler). Second, `tunnel.js` forwards inbound relay messages to `http.request({ hostname: '127.0.0.1', port,... })`, proxying arbitrary requests to services listening on the installer's loopback interface and base64-encoding the responses back to the relay. The port allow-list at `~/.agentlink/tools/proxy-config.json` is itself mutable from the network side via a `proxy_config_update` message (`handleProxyConfigUpdate`), though ports must additionally be marked enabled by the operator. A party who reaches the relay or holds the session URL can execute arbitrary shell commands on the installer's machine and pivot to any localhost service exposed through the tunnel. The remote-development capability is the package's advertised purpose, but the mechanism is a live remote-shell and localhost-tunnel channel to a hardcoded third-party endpoint.