Change the default model
Editmodels.default in librechat.yaml, then upgrade the release with the same command you used to install it. Do not add --reuse-values, so the chart reads the new file:
Rotate the token
Rotate the token before it expires, or at once if you think it leaked. Turn on auto-rotation for production so the token renews before expiry and the owner is notified.1
Regenerate the token
In Access Management → Virtual Accounts, open the account and regenerate its token. Copy the new token now. The console shows it once.
2
Update the cluster secret
Replace the value in place, without printing it:
3
Restart LibreChat
The pod reads the token at start-up, so restart it:Then repeat the completion check on a granted model. Expected:
200.Revoke access
Both actions take effect at once at the gateway:- To stop one model, open Manage Permissions on the virtual account and delete that grant. The trash icon removes it.
- To stop every call from LibreChat, delete the virtual account.
Upgrade safely
The image store and MongoDB useReadWriteOnce volumes. A rolling update can start a new pod that cannot attach a volume while the old pod still holds it. Scale the deployment to zero before the upgrade, then scale back up:
ReadWriteMany.
Remove the release
Uninstall the release, then delete the namespace:Terminating after its resources are gone. A kubernetes finalizer blocks deletion when API discovery fails. Check the finalizers first:
Troubleshooting
Checklist
Use this list to confirm a working setup:- Namespace
librechatand secretlibrechat-credentials-envexist, with four keys - Virtual account created, and its token stored in
librechat-gateway-key - At least one model granted with Provider Integration Access, and its provider-qualified ID noted
-
librechat.yamlsetsmodels.defaultto that ID - Helm install finished, pods show
1/1 Running, and volume claims showBound - Logs show
Connected to MongoDBand the custom endpoint -
/modelsreturns200and lists the granted ID - A completion on the granted ID returns
200 - A completion on a model you did not grant returns
403 - Sign-in works, and a chat shows the granted model in the header badge