How to use other providers/model like google/gemini-3-pro

[nippyin]

Is there a set of instructions on how to use different providers/models?

[Bingtagui404]

Hi, this is a great question. I noticed the docs don't currently cover how to configure alternative LLM providers (like Google Gemini).

Would the maintainers welcome a documentation PR that adds a guide for configuring third-party providers? I'd be happy to contribute one if so.

[alannguyen1]

I ran into this while wiring Gemini-backed web search inside a NemoClaw sandbox on macOS. It turns out there are a couple of practical blockers beyond just missing docs: the default sandbox policy did not allow generativelanguage.googleapis.com, the relevant policy binary allowlists needed /usr/local/bin/node, and OpenClaw's trusted env proxy path still performed a local DNS lookup before creating EnvHttpProxyAgent, which caused getaddrinfo EAI_AGAIN inside the sandbox. I opened a dedicated issue with repro steps, diagnosis, and the local workaround here: #396\n\nIf maintainers prefer, I can also turn the docs part into a PR once there is clarity on the intended fix path.

[andy-ratsirarson]

We hit the same wall trying to use Gemini and dug into the root cause.

The problem: OpenClaw sends store: true in API requests — an OpenAI-specific parameter that non-NVIDIA endpoints like Gemini reject with 400 status code (no body).

The fix: Set supportsStore: false in the model compat inside openclaw.json (baked at Docker build time in the Dockerfile).

Rather than adding Gemini as a named provider, we went with a generic "Custom OpenAI-compatible endpoint" option in the onboarding wizard. Any provider with a /v1/chat/completions endpoint works:

• Google AI Studio (Gemini)
• OpenRouter
• Together AI
• LiteLLM
• Any OpenAI-compatible proxy

The custom provider uses the same gateway-routed architecture as existing providers, so no egress policy or sandbox changes are needed.

PR: #624

[miyoungc]

Closing — this is now fully documented under the Inference section of the docs:

• Inference Options — all supported providers (NVIDIA Endpoints, OpenAI, Anthropic, Google Gemini, OpenAI-compatible endpoints, Anthropic-compatible endpoints, local Ollama) with onboarding walkthroughs.
• Switch Inference Models — change the active model at runtime without re-onboarding.
• Use a Local Inference Server — Ollama, vLLM, NIM, and compatible-endpoint setup.

For Google Gemini specifically, select option 6 during nemoclaw onboard and provide your GEMINI_API_KEY. For any other OpenAI-compatible provider, select option 3 and enter the base URL and model name.