Garry's Opinionated OpenClaw/Hermes Agent Brain
Google Gemini
GBrain calls Gemini natively for chat, query expansion and embeddings: model
strings google:<model>, such as google:gemini-2.5-flash for chat and
google:gemini-embedding-2 for embeddings. Embedding models, dimensions and
pricing are in embedding providers.
Key setup
Set GOOGLE_GENERATIVE_AI_API_KEY, an AI Studio API key. GEMINI_API_KEY works as an alias
when the first variable is unset. You can also set google_api_key in ~/.gbrain/config.json; a
real value in the environment wins over the config file.
Base URL override
GOOGLE_GENERATIVE_AI_BASE_URL sends every native Gemini request to another
endpoint: chat, query expansion and embeddings. Use it for a metering or
logging proxy, such as an eval harness's spend proxy, or for a regional gateway.
export GOOGLE_GENERATIVE_AI_BASE_URL=http://127.0.0.1:8787
- A URL with no version segment gets
/v1betaappended, matching the defaulthttps://generativelanguage.googleapis.com/v1beta. Requests then go to<base>/v1beta/models/<model>:generateContent(chat and expansion) and:embedContent/:batchEmbedContents(embeddings). A proxy that forwards the path upstream works unchanged. - A URL that already ends in a version (
/v1,/v1beta,/v1alpha) is used as written. Trailing slashes are trimmed. - The API key still goes in the
x-goog-api-keyheader, so the endpoint receives your key. Point it only at a proxy you control. - Unset or blank: the SDK default is used, so nothing changes.
The file-plane equivalent is provider_base_urls.google in
~/.gbrain/config.json. The environment variable wins over it. A value merged
from a mounted brain's database never sets this URL, so a shared brain cannot
redirect your key; the same rule covers ANTHROPIC_BASE_URL and
OPENAI_BASE_URL.
Verify the override with one short request through the proxy, for example
gbrain think "ping" --model google:gemini-2.5-flash, and check that the proxy
logged a /v1beta/models/gemini-2.5-flash:generateContent call.