How do you run Gemma 4 in Claude Code?

Claude Code sends every request in Anthropic's Messages format. It reaches another server when ANTHROPIC_BASE_URL is set to that server's address.

Ollama answers in that format, on your own machine and on its cloud. OpenRouter answers in it too.

Claude Code edits files and runs commands through tool calls. Google built Gemma 4 with native function calling, so it can drive those tools.

Google publishes Gemma 4 as open weights under the Apache 2.0 license. Anyone may run it, on their own hardware or on a host's.

Install Ollama and Claude Code, then start Claude Code on Gemma 4:

A bare gemma4 is Ollama's default tag, the small E4B model with a 128K window.

For a larger repository, Ollama says to set the context length to 64K or higher.

The model then runs on your own hardware, and Ollama's plans put no limit on local use. How fast it answers depends on that hardware.

For one developer with a capable machine, this is the simpler answer. A gateway adds nothing there.

Gemma 4 comes in five sizes. Ollama names each one as a tag:

The 26B model runs almost as fast as a 4B model, Google says, because only 4B of its parameters work on each token.

The 31B model is a download of about 20GB. Ollama's cloud runs it without one.

With the Ollama app, name the cloud tag:

Without the app, create an API key on ollama.com. Set ANTHROPIC_BASE_URL to https://ollama.com and put the key in ANTHROPIC_AUTH_TOKEN. Then name the model as Ollama's cloud lists it, gemma4:31b.

Your code then leaves your machine for Ollama's servers. The usage counts against your Ollama plan.

On a team, that is one Ollama key and one set of variables per laptop, with no shared view of who used what. Why a key on every laptop is the expensive part is in keeping API keys off laptops.

Gemma 4 is a model Selan lists, at Ollama as gemma4:31b and at OpenRouter as google/gemma-4-31b-it. An owner connects the account once, then switches Gemma 4 on under Settings → Models.

Developers install the selan CLI, sign in as themselves and name the model:

The part before the first colon is the provider. Everything after it is that provider's own model id, passed through as it is.

With --model, Selan tells Claude Code the model's real context window. Selan lists Gemma 4 at 262,144 tokens, so a session compacts there and not at 200K. If an owner set a lower context limit for your company, it compacts at that instead.

To skip the flag, an owner adds Gemma 4 to Claude Code's picker under Settings → Model pickers. It then appears in /model beside the Claude models. The details are in Run any model.

If that account is out of quota, the request fails rather than moving to the other one.

Selan reaches Ollama's cloud at ollama.com, never an Ollama server on a laptop. How Ollama itself runs in Claude Code is in using Ollama with Claude Code. Every provider is on Providers and models.

• One usage record. Gemma 4 turns land beside Claude turns, per person and per repository. Usage per developer shows what an owner sees.
• One spend limit. An OpenRouter turn is metered, so it counts against a member's limit. A turn on an Ollama plan does not. See spend limits.
• No Ollama or OpenRouter key on a laptop. Keys stay on the gateway. A developer holds a Selan session you can revoke.
• Codex on the same plan. selan codex --model ollama:gemma4:31b uses Ollama's own Responses API. See Codex.
• Cheap agent work. Selan's comment-cleaning agent names openrouter:google/gemma-4-31b-it as its model, because the work is mechanical. See the minimise-comments agent.

• Local turns are not on the record. A developer running Gemma 4 on a laptop by Ollama's own setup bypasses Selan. Its usage record and spend limits do not see those turns.
• The size is part of the answer. Selan runs the 31B model. A teammate comparing it with E4B on a laptop is comparing two different models.
• The model must be switched on. A model your company cannot use is refused before Claude Code starts, not at the first request.

Free for up to three seats, and the Free plan needs no card.