Google

Google is natively supported in fast-agent using the Google GenAI libraries.

Google models have support for attaching YouTube URLs for video and transcript understanding. Gemini models that support Grounding with Google Search can also use provider-side web search via the standard web_search model-string option.

YAML Configuration:

google:
  api_key: "your_google_key"

Environment Variables:

• GOOGLE_API_KEY: Your Google API key

Reasoning, search, and multimodal input

Google models support model-dependent thinking controls, structured outputs, multimodal inputs, and Grounding with Google Search.

fast-agent --model gemini
fast-agent --model "gemini37flash?reasoning=medium"
fast-agent --model "gemini3?reasoning=auto"
fast-agent --model "google.gemini-3.1-pro-preview?reasoning=high"
fast-agent --model "gemini3?web_search=on"
fast-agent --model "google.gemini-3.7-flash?web_search=on"
fast-agent --model "gemini37flash?service_tier=flex"

Useful query parameters:

• reasoning=auto|minimal|low|medium|high|off where the selected Gemini model advertises thinking controls
• web_search=on|off for Grounding with Google Search on supported Gemini models
• structured=json for JSON schema structured outputs
• sampling controls such as temperature, top_p, and top_k where applicable

Gemini 3.7 Flash is a GA coding and agent model with a 1M-token context window, 64K maximum output, and low, medium, or high thinking levels (medium by default). It does not accept sampling parameters, numeric thinking budgets, or candidate_count; fast-agent omits those fields from native requests. Introductory pricing through December 31, 2026 is $0.75 per million input tokens and $3.75 per million output tokens.

Flex inference

Gemini 3.7 Flash supports Google's synchronous, best-effort Flex tier:

fast-agent --model "gemini37flash?service_tier=flex"

Or configure Flex as the Google default:

google:
  api_key: "your_google_key"
  service_tier: flex

In the interactive prompt, press Shift+Tab to cycle between Standard and Flex. The existing /fast flex command is a shortcut for /model fast flex; /fast off returns to Standard. The command name predates Gemini support, so Flex remains available even though it is intentionally slower than the default tier.

Flex targets latency-tolerant work and may take 1–15 minutes. Fast-agent uses a 900-second stream/request timeout by default while Flex is selected; an explicit streaming_timeout query still takes precedence. Capacity errors such as HTTP 503 remain retryable.

Developer API requests use Google's native serviceTier: flex field. Vertex Flex requires the global location and uses the documented Vertex shared/Flex routing headers with API version v1.

When web_search=on, fast-agent sends Google's native GoogleSearch tool and formats grounding metadata as citations in the text response when Google returns citation spans. Search can be toggled from the /model web_search command or the interactive model controls on models that advertise support.

Current fast-agent metadata marks Google Search as supported on Gemini 2.0 Flash, Gemini 2.5, and Gemini 3.x aliases, including Gemini 3.7 Flash. Check Models Reference for the generated capability view.

Model Name Aliases:

OpenAI Mode

You can also access Google via the OpenAI-compatible provider. Use googleoai in the YAML file, or GOOGLEOAI_API_KEY for API key access; that path uses the Google OpenAI-compatible endpoint by default.