Claude Code with other models — a gateway, or a model server on your Mac

Claude Code talks to Anthropic by default: this Mac's Claude login, a setup token or an Anthropic API key, exactly as in Agents and accounts. It can also get its model from:

• a gateway — OpenRouter: one key, models from many makers, through the gateway's Anthropic-compatible endpoint;
• a model server on this Mac — Ollama, LM Studio, llama.cpp's llama-server, LiteLLM, vLLM or oMLX, reached by its port. No key.

Anthropic supports Claude models only. Another maker's model through a gateway or your own server is yours to judge — it may not handle Claude Code's tools as well.

Through a gateway (OpenRouter)#

Add the gateway's key once — it stays in this Mac's keychain, like every key:

doz account add my-or --provider openrouter

Then try Claude Code through it, with a model by the gateway's name for it:

doz try claude-code --gateway openrouter --model anthropic/claude-sonnet-4.6

doz models --gateway openrouter lists the gateway's models. The sandbox gets the gateway's address and a placeholder in ANTHROPIC_AUTH_TOKEN; Dozable's proxy on the Mac swaps in the real key, and only on the way to openrouter.ai. A placeholder sent anywhere else is refused. A Claude subscription is never used this way: the gateway key replaces it.

A model server on your Mac#

Start the server on your Mac first (ollama serve, LM Studio's server, …), then:

doz try claude-code --provider local --local-port 11434 --model qwen3-coder

• --local-port is the port the server listens on (Ollama 11434, LM Studio 1234, llama-server 8080, LiteLLM 4000, vLLM and oMLX 8000).
• --model is the server's own name for the model (ollama list, LM Studio's list). Every alias Claude Code uses (and its subagents) gets that model, so it never asks your server for a Claude model it does not have.
• There is no key: Claude Code is given a fixed word in ANTHROPIC_AUTH_TOKEN, which the local servers ignore.

doz models --provider local lists the servers known to work, with the port each uses by default.

The servers must speak Anthropic's Messages API (/v1/messages) — Ollama (0.14 or later), LM Studio (0.4.1 or later), llama-server, LiteLLM, vLLM and oMLX do. A server that speaks only OpenAI's chat API (mlx_lm.server) needs a translator in between: not yet.

In a project file:

model_source: local
local_port: 11434
model: qwen3-coder

What the sandbox can reach on your Mac#

A sandbox has no network of its own: everything it sends goes through Dozable's proxy, which runs on your Mac. So:

• One name, one port. The sandbox reaches your server as http://mac.dozable.internal:PORT, for the port it was made with — the proxy sends that to 127.0.0.1:PORT on your Mac. Any other port of that name is refused, and a sandbox made without a port cannot resolve the name at all.
• Nothing else on your Mac's localhost. 127.0.0.1, localhost and ::1 through the proxy are refused for every sandbox, whatever its network — Open included. (Inside the sandbox, localhost is the sandbox's own.)
• It is not a network permission: no preset or edit widens it. The sandbox's Network tab says it ("This Mac: port 11434 only"), and so does the agent's own description of its sandbox.
• doz doctor checks that something answers on each such port.

The port is chosen when the sandbox is made; for another port, make another sandbox.

In the dashboard#

Try (Claude Code's setup) and New sandbox (its Account step) show Where Claude Code gets its model: Anthropic with its usual account choice, OpenRouter with its key, and A model server on this Mac with a port (one click for the known servers) and the model. The sandbox's page shows where its model comes from ("via OpenRouter · model …", "via this Mac's port 11434 · model …"). Accounts & keys says a server on your Mac needs no key.