Google Gemini

Access Gemini models via the Google Generative Language API with support for API keys, Gemini CLI OAuth tokens, and long context windows up to 2M tokens.

Prerequisites

• A Google AI Studio API key from aistudio.google.com, or
• Gemini CLI installed and authenticated (gemini command), or
• A GEMINI_API_KEY or GOOGLE_API_KEY environment variable

Quick Setup

1. Get API Key

Option A: API Key (recommended for most users)

1. Visit aistudio.google.com/app/apikey
2. Click Create API key
3. Copy the key

Option B: Gemini CLI (zero-config for existing users)

If you already use the Gemini CLI, PRX automatically detects your OAuth token from ~/.gemini/oauth_creds.json. No additional configuration needed.

2. Configure

[default]
provider = "gemini"
model = "gemini-2.5-flash"

[providers.gemini]
api_key = "${GEMINI_API_KEY}"

Or set the environment variable:

export GEMINI_API_KEY="AIza..."

3. Verify

prx doctor models

Available Models

Model	Context	Vision	Tool Use	Notes
gemini-2.5-pro	1M	Yes	Yes	Most capable Gemini model
gemini-2.5-flash	1M	Yes	Yes	Fast and cost-effective
gemini-2.0-flash	1M	Yes	Yes	Previous generation flash
gemini-1.5-pro	2M	Yes	Yes	Longest context window
gemini-1.5-flash	1M	Yes	Yes	Previous generation

Configuration Reference

Field	Type	Default	Description
api_key	string	optional	Google AI API key (AIza...)
model	string	gemini-2.5-flash	Default model to use

Features

Multiple Authentication Methods

PRX resolves Gemini credentials in this priority order:

Priority	Source	How it works
1	Explicit API key in config	Sent as ?key= query parameter to public API
2	GEMINI_API_KEY env var	Same as above
3	GOOGLE_API_KEY env var	Same as above
4	Gemini CLI OAuth token	Sent as Authorization: Bearer to internal Code Assist API

Gemini CLI OAuth Integration

If you have authenticated with the Gemini CLI (gemini command), PRX automatically:

1. Reads ~/.gemini/oauth_creds.json
2. Checks token expiry (skips expired tokens with a warning)
3. Routes requests to Google's internal Code Assist API (cloudcode-pa.googleapis.com) using the proper envelope format

This means existing Gemini CLI users can use PRX with zero additional setup.

Long Context Windows

Gemini models support extremely long context windows (up to 2M tokens for Gemini 1.5 Pro). PRX sets maxOutputTokens to 8192 by default. The full conversation history is sent as contents with proper role mapping (user/model).

System Instructions

System prompts are sent using Gemini's native systemInstruction field (not as a regular message), ensuring they are handled correctly by the model.

Automatic Model Name Formatting

PRX automatically prepends models/ to model names when needed. Both gemini-2.5-flash and models/gemini-2.5-flash work correctly.

Provider Aliases

The following names all resolve to the Gemini provider:

• gemini
• google
• google-gemini

Troubleshooting

"Gemini API key not found"

PRX could not find any authentication. Options:

1. Set GEMINI_API_KEY environment variable
2. Run the gemini CLI to authenticate (tokens will be reused automatically)
3. Get an API key from aistudio.google.com/app/apikey
4. Run prx onboard to configure interactively

"400 Bad Request: API key not valid" with Gemini CLI

This occurs when OAuth tokens from the Gemini CLI are sent to the public API endpoint. PRX handles this by routing OAuth tokens to the internal cloudcode-pa.googleapis.com endpoint automatically. If you see this error, ensure you are using the latest version of PRX.

"Gemini CLI OAuth token expired"

Re-run the gemini CLI to refresh your token. PRX does not refresh Gemini CLI tokens automatically (unlike Anthropic OAuth tokens).

403 Forbidden

Your API key may not have the Generative Language API enabled. Go to the Google Cloud Console and enable the Generative Language API for your project.