Gemini CLI and Google GenAI

Gemini CLI and the programs built on Google’s GenAI SDKs use gateway mode: the local daemon masks requests and restores responses. It covers the Gemini API with an API key, and Gemini on Vertex AI, with a Google Cloud project or in express mode with an API key.

Base URLs

Service	Instead of	Use	Variable
Gemini API	https://generativelanguage.googleapis.com	http://127.0.0.1:5283/gemini	GOOGLE_GEMINI_BASE_URL
Vertex AI	https://<location>-aiplatform.googleapis.com	http://127.0.0.1:5283/vertex	GOOGLE_VERTEX_BASE_URL

Keep your API key or your Google Cloud login where it is. The gateway forwards it unchanged (a key in the x-goog-api-key header or in the address’s key parameter, a Vertex token in Authorization) and never writes its value to logs, to the dashboard or to a capture. On Vertex, each request goes to the address of the location it names.

Gemini CLI

Run Gemini CLI through Latet with the wrapper. It starts the daemon if needed, sets both base URLs for that run only, and passes every argument to gemini:

latet gemini
latet gemini -p "write the release notes for the payments module"

latet run gemini does the same. Gemini CLI takes its address only from the environment, so it is not available in global mode: latet global-install does not set it up, and there is no latet agents apply gemini. A change for good would not reach a project with an .env file of its own, and Gemini CLI would then talk to Google directly without saying so.

Choose an API key or Vertex AI

In Gemini CLI, /auth chooses how it reaches its model, and the choice is kept in ~/.gemini/settings.json:

~/.gemini/settings.json

{ "security": { "auth": { "selectedType": "gemini-api-key" } } }

• An API key (gemini-api-key): GEMINI_API_KEY holds it.
• Vertex AI (vertex-ai): GOOGLE_GENAI_USE_VERTEXAI=true, with GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION and your Google Cloud login, or GOOGLE_API_KEY alone in express mode.

When nothing is chosen yet, latet gemini offers the API key in Gemini CLI’s dialog, or Vertex AI when GOOGLE_GENAI_USE_VERTEXAI=true is set. A run with -p needs the choice made first: set selectedType as above.

Signed in with Google, whether your settings, the project’s (.gemini/settings.json in the folder you start it in) or the machine’s choose it, Gemini CLI sends its requests to an endpoint Latet does not read, so nothing would be masked. latet gemini does not start it then, and says where the choice is made and what to choose instead.

Usage statistics

Gemini CLI sends its usage statistics to Google directly, not through Latet. To turn them off:

~/.gemini/settings.json

{ "privacy": { "usageStatisticsEnabled": false } }

Scripts and the Google GenAI SDKs

latet run sets GOOGLE_GEMINI_BASE_URL and GOOGLE_VERTEX_BASE_URL beside the OpenAI and Anthropic variables, and latet env prints them for a whole shell. Google’s SDKs for Python (google-genai) and JavaScript (@google/genai) read them:

welcome.py

from google import genai

client = genai.Client()  # reads GEMINI_API_KEY and GOOGLE_GEMINI_BASE_URL
reply = client.models.generate_content(
    model="gemini-3-pro",
    contents="Write a short welcome email to emily.carter@northwind-legal.com",
)
print(reply.text)

Run it with latet run python welcome.py: Google reads a fake address of the same shape, and the script prints the email with the real one. Streamed answers work the same way, chunk by chunk.

Google’s OpenAI-compatible endpoint goes through Latet too: give an OpenAI client the base URL http://127.0.0.1:5283/gemini/v1beta/openai.

What the gateway reads

• Generating, streamed or not (generateContent, streamGenerateContent), counting tokens (countTokens), embeddings (embedContent, batchEmbedContents) and creating a cache (cachedContents): the system instruction, every message, function calls and their results, the functions’ declarations and the schema of the answer are masked, and the answers are restored.
• Images sent in the request are read and masked with the optional text recognition. Other files, and files given by their address, are sent as they are and flagged Not inspected in the dashboard.
• The model’s thinking, and the code Google runs for it, are left as the model wrote them.
• Any other request with a body on these routes is refused, with an error that says so, and a WebSocket of the Live API is closed before anything goes through: the Interactions API, the Live API, the Batch API, file uploads and Vertex’s :predict (its own embeddings and image models) carry a conversation in forms Latet does not read yet (they are on the roadmap). A request without a body, such as the list of models, passes.

Limits

• Gemini CLI signed in with Google is not covered (see above).
• Antigravity CLI takes a base URL with an API key (GOOGLE_GEMINI_BASE_URL), but its documentation does not say which of Google’s protocols it then speaks, and it is not verified with Latet. If it uses the Interactions API, Latet refuses its requests rather than send them unmasked.
• An answer that uses an explicit cache gets back the originals of the values its own conversation carries.