Google is a first-party model provider — it
produces the Gemini family and serves them from its own API. It is the
zero-config default for the gemini llm tool, and a first-class
model provider: once connected, operator lists its available models live
for delegators.
Model provider ≠ llm tool. Google (the provider) serves the models; Gemini CLI (the llm tool) is the CLI. A delegator pairs a tool with a provider’s model.
Connect
Operator references your key by env-var name — it never stores the secret:
export GEMINI_API_KEY="..."
A provider is connected when its /models probe succeeds. Google then shows
● connected in the Model Providers view with its live model list.
Listing models
Operator probes https://generativelanguage.googleapis.com/v1beta/models
(model list reference) and lists whatever the
API returns, stripping the models/ prefix so the id matches --model:
GET /api/v1/model-servers/kinds/google-api/models # { reachable, models[], error? }
Use from a delegator
[[delegators]]
name = "gemini-pro"
llm_tool = "gemini"
model = "gemini-1.5-pro"
# model_server omitted → resolves to the implicit google-api default
Custom endpoint / proxy
Declare a google-api server with an explicit base_url to point Gemini at a
proxy; that base URL is injected at spawn (GOOGLE_GEMINI_BASE_URL). The probe
default is probe-only and never changes the launch path on its own.