Hi everyone,Sharing this small tool called os_sup_mcp.It’s an OTP-style supervisor that spins up isolated child BEAM nodes (full Elixir apps) and exposes them via a simple MCP interface (JSON-RPC over TCP).The cool part? AI coding agents (like Claude Code, Cursor, etc.) can now:
-
Start/stop/restart nodes
-
Run eval on live instances with full access to real processes, ETS tables, supervisors, and app state
-
Debug complex stuff (race conditions, GenServers, etc.) in seconds instead of the usual “edit → restart → test” loop
-
Solve the pain of running eval MCP directly on projects, where agents like Claude Code don’t reconnect by themselves and the protocol isn’t designed to handle failures robustly—isolated nodes make it reliable
It’s early but already feels useful for agentic workflows.
Repo:
Would love any feedback or ideas!
Thanks!
