Scoped MCP, not a pasted terminal
Giving an AI editor a raw shell is convenient until it runs kill -9 on the wrong pid. Rig's local MCP hub is the other contract: scoped tools against the processes you already named in the app.
The hub listens on localhost. You pick full, control, or read. Rotating the token unlinks every editor at once. The agent can list projects, read logs, check health, see who owns a port, diagnose a crash, and start or stop a process. It cannot browse your disk as if it were you.
Why this is the product
Agents guess at runtime. They paste a stack trace from ten minutes ago. They assume Postgres is up. Rig's hub is how they see what is actually running on this Mac, now. That is the difference between "the checkout 500s" as a vibe and ECONNREFUSED :5432 from read_logs.
We built it this way because the alternative — paste a terminal into the prompt — does not scale past one developer who trusts the model that day.
What we do not do
We do not ship a cloud agent that SSHs into your laptop. We do not ask for the repo. Source stays on disk. The MCP session is a window into process state, not a second filesystem.
How to link it
Install Rig, open Connections, one click for Cursor, Claude Code, Codex, Windsurf, or VS Code. Details: What is MCP?, MCP in Rig, MCP for Cursor.
If you want the marketing version of the same idea, it is on /ai.