Your own model servers

The "Models" tab in the settings connects someone else's server that speaks the Anthropic protocol: llama-server, llama-swap, LM Studio — anything that serves /v1/messages.

The server list lives on the server: added from the laptop, the phone sees it right away.

What you set

The address and the key, on/off, refreshing the model list and a connection check.

A provider's model appears in the common picker as a separate group under the server's name, as a link like llamacpp/gpt-oss-20b. An unloaded one is labelled: the first reply on it will come after the weights are fetched.

There is no separate provider switch — choosing a model is choosing a server.

Two quirks that can't be worked around

Switching to a provider's model restarts the session. Claude Code has no API of its own for "change provider": the connection is the environment of the CLI's child process, and that is set at start.

Such models have no effort levels. effort is a property of Anthropic's models; someone else's server simply doesn't have it.

The connection check

It goes step by step, and every failed step gives a hint:

1. the address answers;
2. the model list;
3. the Anthropic protocol — a 404 means an OpenAI-compatible server;
4. tool schemas — llama.cpp trips over heavy ones, and such a tool is simply switched off for the session;
5. a streamed reply.

On an unloaded model the check takes up to a minute, and that is normal.