Shelby Stacks
Self-hosted integration and chat stacks on the always-on server: Nango for connector auth and proxying, LibreChat as a private multi-model chat, both reachable only on the private network.
Every new integration used to mean another OAuth dance, another secret file and another place tokens could leak. Chat with several models meant several browser tabs and no shared memory.
a dedicated container profile, separate from the production VM
Nango holds the connector authorizations and proxies the calls
LibreChat fronts the models with tools from the knowledge servers
tailnet-only hostnames, admin user, no public ingress
autostart on boot and a nightly backup that refuses to run when the disk is nearly full
- Runs in its own Colima profile so a chat stack can never starve the production services
- Knowledge servers (documentation lookup, memory, knowledge graph) are wired in as tools
- Pinned docker context in the stack script; switching profiles no longer switches the world
- Free self-hosted Nango is auth plus proxy only; that was enough
New integrations become configuration instead of engineering, and the model conversations stay on hardware Aragorn owns.