OpenClaw with Gemini

The Google plugin gives OpenClaw the Gemini models through Google AI Studio, plus image generation, media understanding for images, audio and video, text to speech and web search through Gemini grounding. The provider is google, the auth is an API key, and there is a managed-cloud variant for Vertex AI. Here is the recommended path, what changes the model list without a release, the CLI runtime and the two OAuth paths the docs now rule out, and what free means in this case.

The recommended path

• Create a free key in Google AI Studio; the Gemini and Google key variables are both accepted, so use whichever you already have.
• Run onboarding with the Gemini API-key choice, or pass the key non-interactively with the local mode and the accept-risk flag for a scripted install.
• Set a default model, the docs’ example being a Gemini 3.1 Pro preview, and verify it with the models list command for the provider.
• With a key configured, OpenClaw refreshes the AI Studio text-model catalog from Google’s models API, so new Pro, Flash and Flash-Lite variants appear without waiting for an OpenClaw release; a failed refresh keeps the last inventory, and a successful empty response clears the discovered models.

Vertex AI

Use the Vertex provider when the Gateway already runs inside a managed Google Cloud environment, with Application Default Credentials instead of a key. Vertex keeps a separate static catalog rather than the refreshed AI Studio list.

OpenClaw does not offer new Gemini CLI OAuth or Antigravity OAuth setup.

The CLI runtime

For advanced use only, a canonical Google model can run through an installed Gemini CLI, from Homebrew or npm, while authentication stays on the AI Studio key: the model keeps its Google reference and is opted into the CLI runtime in config. It is not a way to log in with a Google account. Google ended consumer Gemini CLI login on 18 June 2026 and the Antigravity terms prohibit third-party tools from using that service, so the key or Vertex is the path. OpenClaw models and providers covers how references resolve, and OpenClaw web search shows Gemini grounding as a search provider.

What free means

The key is free to create; the docs promise nothing about free inference, and any free-tier limit is Google’s, not OpenClaw’s. For genuinely free serving the docs point elsewhere: OpenClaw and NVIDIA for hosted open models and Free models for OpenClaw for a model on your own hardware, each with its own caveat about smaller models and prompt injection.

On Diali

On Diali the Gemini models are in the managed catalog and paid from included credits, so there is no key to create; a key of your own is the option if you want Google to bill you directly. Hosted OpenClaw on Diali is the assistant.

• A free AI Studio key, one onboarding choice, a default model.
• The catalog refreshes from Google; Vertex keeps its own.
• The CLI runtime is a runtime, not a login.