What leaves the server
Only this, and most rows only when you set them up:A server without internet
Prepare the model and the images on a machine with internet and the same CPU architecture as the server (amd64 keeps the model as safetensors, arm64 as ONNX: a volume filled on the other architecture does not start).1
On a machine with internet: build the image and fill the model
Build the image and run the install once until
/readyz is 200 (the embedder has then downloaded the model). Then export the images and the model volume:2
Transfer the files
Copy the install directory (without
.env and backups/), tb-images.tar.gz and tb-models.tar.gz to the server.3
On the server, before the first start
cpu-arm64-1.9.4 goes instead of cpu-1.9.4.docker compose may warn that the volume was not created by Compose: it uses it anyway.4
Continue with the install
Go to Install from the
.env file. The embedder starts from the local snapshot. The first admin (First sign-in and admin) runs inside the image and needs no internet.- For the monthly restore check (Backup and restore) keep the
docker tagline: the script then uses the loaded image of the release instead of building a new one, and the loaded Node image runs its helper containers. - Sign-in, connectors and TLS from a public certificate authority need their hosts from the table. Without them nobody can sign in to the web app (the page says sign-in is not set up): what remains are the operator commands, the upload and personal tokens (
./tb admin token create) for MCP and the API, and TLS behind your own proxy with your own certificate.
Limits
- Without the Google or Microsoft hosts there is no sign-in with those accounts and no connectors for Drive and SharePoint.
- Without
claude.aiandchatgpt.comCIMD does not work: turn on DCR or pre-register a client, see OAuth. - The embedder for a model change on such a server is filled the same way.