Desktop Model Servers

Model servers let Bike4Mind Desktop chat with models that run on another machine, such as a GPU box at home or in a lab. You add the server once under Settings > Model servers, and its models appear in the model picker next to the ones on your own computer.

Desktop supports two kinds of server:

• Ollama.
• OpenAI-compatible servers such as llama-server, LM Studio and vLLM.

Each server has an Access setting. Chat only lets Desktop use the server's models for chat. Chat and manage marks the server as one you administer. Desktop does not pull or remove models on a remote server with either setting. Chat only is the default, and the setting applies to this device only.

When to use this

Use a model server when the model is too large for your computer, or when several people or devices should share one machine's GPU. If the model runs on the same computer as Desktop, you do not need this: Desktop manages local runtimes itself.

How Desktop protects your credentials

A server can need a bearer token or a custom header (for example X-Api-Key). Desktop applies these rules:

• The token or header value is encrypted with your operating system's secure storage and is never shown again after you save it.
• Credentials are sent only to the address you saved, never to any other host.
• Desktop does not follow redirects. A server that redirects, including http to https, shows an error and no credentials are sent on. Enter the final address instead.
• Credentials travel only over https, or over plain http to this computer (loopback) or a Tailscale IPv6 address (fd7a:115c:a1e0::/48).
• Plain http to a private network address (192.168.x.x, 10.x.x.x, 172.16.x.x to 172.31.x.x, Tailscale IPv4 100.64.0.0/10, which covers 100.64.x.x to 100.127.x.x, and IPv6 unique local addresses in fc00::/7) works only without credentials. Desktop marks these servers Unencrypted.
• A public address needs https with a valid certificate. Plain http to a public address is refused. Link-local addresses (169.254.x.x, fe80::/10) count as public.
• A host name is judged by the addresses it resolves to, and the least trusted one decides. localhost counts as this computer.

If your computer has no secure storage available, Desktop keeps the credential only until you quit and marks the server Credentials kept until Desktop quits.

Pick a setup

Setup	Address you enter	Credentials
Tailscale with HTTPS	https://gpu.tailnet-name.ts.net	Optional
Caddy in front of Ollama	https://ollama.example.com	Bearer token
SSH local port forward	http://127.0.0.1:21434	None needed
Plain LAN	http://192.168.1.20:11434	None, shown as Unencrypted

Prepare the server

Ollama listens only on 127.0.0.1 by default. For the proxy and SSH setups below, leave it that way: the proxy or the SSH daemon is the only thing that needs to reach it. Set OLLAMA_HOST only when another machine must connect to Ollama directly, as in the LAN setup:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Never expose Ollama on a public address without a proxy that requires a token. Ollama has no authentication of its own, so anyone who finds the address can run your models and pull or delete them. Desktop warns you when a public address answers without credentials.

For llama-server, start it with --host and --port the same way, and --api-key if you want it to check a bearer token itself.

Tailscale

Tailscale gives each machine a private address that only your tailnet can reach. Install Tailscale on both machines and sign in to the same tailnet.

Desktop trusts plain http with credentials only on the Tailscale IPv6 address. The IPv4 address in 100.64.0.0/10 counts as a private network address, so it works without credentials and is marked Unencrypted. To use a token or to avoid the Unencrypted mark, serve the model over HTTPS with Tailscale's own certificates:

tailscale serve --bg --https=443 http://127.0.0.1:11434

Then add https://<machine>.<tailnet>.ts.net in Settings > Model servers. Check the Tailscale documentation for the current serve syntax and to enable HTTPS certificates for your tailnet.

Ollama rejects requests whose Host header is not a local name, and Tailscale's documentation does not say whether serve rewrites that header. Before you add the server, run curl -i https://<machine>.<tailnet>.ts.net/api/version from another machine on the tailnet. If it returns 403, use the Caddy setup below on the same machine instead, or connect over the Tailscale IPv6 address without serve.

Caddy with https and a bearer token

Use this when the server has a public name and you want a token in front of Ollama. Caddy gets a certificate for the name automatically. Point the DNS record for the name at the server and open ports 80 and 443.

Create a Caddyfile:

{$SITE_ADDRESS} {
	@unauthorized not header Authorization "Bearer {$OLLAMA_TOKEN}"
	respond @unauthorized "Unauthorized" 401
	reverse_proxy 127.0.0.1:11434 {
		header_up Host {upstream_hostport}
	}
}

The header_up line is required. Caddy passes the public name through as the Host header, and Ollama answers 403 to any Host that is not a local name.

Generate a token and run Caddy with both values set:

export OLLAMA_TOKEN="$(openssl rand -hex 32)"
export SITE_ADDRESS=ollama.example.com
caddy run --config Caddyfile

Keep the token out of the Caddyfile itself so you do not commit it by accident. In Desktop, add https://ollama.example.com with Credentials set to Bearer token and paste the token.

You can confirm the proxy before you add it:

curl -i https://ollama.example.com/api/version
curl -i -H "Authorization: Bearer $OLLAMA_TOKEN" https://ollama.example.com/api/version

The first call returns 401. The second returns the Ollama version.

SSH local port forward

If you can already SSH to the server, forward its Ollama port to your computer. Desktop then sees a loopback address, which it trusts without https, and SSH encrypts the traffic.

ssh -N -L 21434:127.0.0.1:11434 user@gpu.example.com

Use a local port other than 11434. Desktop's This computer entry always uses 127.0.0.1:11434, so a tunnel on that port would clash with a local Ollama or make This computer talk to the remote machine.

Add http://127.0.0.1:21434 in Settings > Model servers. The server shows Unreachable whenever the SSH session is closed, so keep it running, or use a tool such as autossh to restart it.

Plain LAN without credentials

On a trusted home or office network you can connect directly:

1. On the server, run Ollama with OLLAMA_HOST=0.0.0.0:11434.
2. In Desktop, add http://192.168.1.20:11434 with Credentials set to None.

Desktop marks the server Unencrypted. Anyone on that network can read your conversations and the replies. Use this only on a network you trust, and do not use it on shared or public Wi-Fi. If you need a token, use one of the setups above.

Add a server in Desktop

1. Open Settings > Model servers.
2. Under Add a server, enter the address. A bare host:port is treated as http.
3. Set Credentials to None, Bearer token or Custom header, and fill in the token, or the header name and value.
4. Select Test. Desktop reports the kind, version, model count and latency, or the reason it could not connect.
5. After a successful Test, Desktop shows the Name, Kind and Access fields. Kind (Ollama or OpenAI-compatible) is filled in from the Test result. Check the fields, set Access (Chat only or Chat and manage), then select Save server.

The address must not include a username, password, query string or fragment. If the server sits behind a proxy path, include the path.

If Test warns that a public address answered without credentials, put a proxy with a token in front of it. Saving anyway requires you to tick Save it anyway.

Troubleshooting

• Needs credentials: the server returned 401 or 403, or Desktop no longer has the saved credential. Check the token, or the header name and value. If the message says the saved credential could not be read, the secure storage lost it, for example after a keyring reset. If it says the credential was kept only until Desktop quit, it was session-only and Desktop restarted. In both cases open the server's credentials and enter the value again.
• Needs credentials with a correct token, behind a proxy: Ollama may be answering 403 because the Host header was not rewritten. Check the Caddy setup above, and test with curl -i against the address.
• Unreachable: the address, port or firewall is wrong, or the SSH forward is closed.
• Refused: the server redirected, its certificate is not trusted, or the address is not allowed. Read the message under Test. If you were redirected to another address, enter the final https address. If the redirect stayed on the same address, check the path you entered.
• The certificate is not trusted: Test shows this message, and the server shows Refused, when the certificate fails verification. Use a certificate from a trusted authority or a local CA your computer already trusts. Tailscale certificates also work.