Google Gemini¶

Autobot features native, highly optimized support for Google Gemini LLMs. It directly implements the generateContent and cachedContents REST APIs, offering native context caching to drastically reduce token usage and cost for long-running autonomous tasks.

Setup¶

1. Get an API key¶

Create an API key at aistudio.google.com.

2. Configure credentials¶

Add your API key to the .env file:

GEMINI_API_KEY=...

Or use the interactive setup:

autobot setup
# Select "Google Gemini" as provider

3. Configure the provider¶

In config.yml:

agents:
  defaults:
    model: "gemini/gemini-3.5-flash"

providers:
  gemini:
    api_key: "${GEMINI_API_KEY}"

4. Verify¶

autobot doctor
# Should show: ✓ LLM provider configured (gemini)

Model naming¶

Models use the gemini/ prefix followed by the Google model ID:

# Gemini 3.5
model: "gemini/gemini-3.5-pro"
model: "gemini/gemini-3.5-flash"

# Gemini 2.5
model: "gemini/gemini-2.5-pro"
model: "gemini/gemini-2.5-flash"

The gemini/ prefix tells autobot to route to the Gemini API. It is stripped before sending to the API.

See the full model list in the Gemini docs.

Native Context Caching¶

Autobot automatically leverages Gemini's Context Caching API to optimize costs for long-running or looping agents.

If your agent's system prompt and tools payload exceeds 8,000 characters, Autobot will: 1. Hash the system state. 2. Explicitly cache it on Google's servers for 1 hour (ttl: "3600s"). 3. Reuse that cache on subsequent loops or prompts.

This is highly recommended for autonomous agent execution as it routinely drops token costs by over 90%.

Configuration reference¶

Field	Required	Default	Description
`api_key`	Yes	—	Google AI API key
`api_base`	No	`https://generativelanguage.googleapis.com/v1beta`	Custom API endpoint

Troubleshooting¶

Enable debug logging to see request/response details and cache usage:

LOG_LEVEL=DEBUG autobot agent -m "Hello"

Look for:

Created explicit Gemini cache: ... — confirms caching is working and saving tokens
HTTP 4xx/5xx: ... — API errors with details

Common issues¶

"HTTP 400: Function call is missing a thought_signature" — This indicates an outdated provider version that is failing to pass back Gemini's internal reasoning state. Update autobot to the latest version.

"API error: API key not valid" — Invalid or expired API key. Verify at aistudio.google.com.

"API error: Resource has been exhausted" — Rate limit or quota exceeded. Check your usage and limits in Google AI Studio.