openai-gemini

To use the proxy, point your software to your deployed API address and provide your Gemini API key. The API base should typically follow this format: https://your-proxy-address.vercel.app/v1

For command-line tools, you may need to set one of the following environment variables:

• OPENAI_BASE_URL
• OPENAI_API_BASE

OPENAI_BASE_URL="https://my-super-proxy.vercel.app/v1"
# or
OPENAI_API_BASE="https://my-super-proxy.vercel.app/v1"

You can run the proxy locally using Node, Deno, or Bun.

Node.js

1. Install dependencies: npm install.
2. Start the server: npm run start.
3. For development mode (with watch): npm install --include=dev then npm run dev.

Deno and Bun

• Start the server: npm run start:deno or npm run start:bun.
• Development mode: npm run dev:deno or npm run dev:bun.

# Node.js example
npm install
npm run start

You can trigger Google Search capabilities by using specific model suffixes or including the tool in your request.

1. Model Suffix: Use a model name ending in :search (e.g., gemini-1.5-flash:search).
2. Preview Models: Models containing -search-preview in their name will automatically have the googleSearch tool enabled.
3. Explicit Tools: Pass tools in the request body containing { "googleSearch": {} }.

To use the built-in web search tool, append :search to the model name.

Example: gemini-2.5-flash:search

Features unique to Gemini that are not part of the standard OpenAI spec can be enabled using the extra_body field in your request.

One notable example is thinking_config. For a full list of supported extra fields, refer to the Gemini API docs.

When calling POST /chat/completions, you can pass Google-specific configurations using the extra_body.google field in your request. This allows access to features not standard in the OpenAI API.

Supported extra_body.google keys:

• safety_settings: Configure content filtering thresholds.
• cached_content: Use context caching.
• thinking_config: Configure reasoning/thinking behavior (for supported models).

Requests will use the specified model if the name starts with gemini-, gemma-, or models/. If no model is specified, the following defaults apply:

• chat/completions: gemini-flash-latest
• embeddings: gemini-embedding-001

The chat/completions endpoint implements most parameters applicable to both APIs:

• messages (including content, role [system, user, assistant, tool], and tool_calls)
• model
• frequency_penalty
• max_tokens, max_completion_tokens
• n (supports candidateCount < 8, not for streaming)
• presence_penalty
• reasoning_effort
• response_format (json_object, json_schema, text)
• seed
• stop (string or array)
• stream and stream_options (including include_usage)
• temperature (Gemini supports up to infinity)
• top_p
• tools and tool_choice
• extra_body (for Gemini-specific features)

Not implemented: logit_bias, logprobs, top_logprobs, parallel_tool_calls (always active in Gemini), and completions endpoint.

The embeddings endpoint supports:

• dimensions

Generate vector embeddings for input text.

Request Parameters:

• model (string): The model to use (e.g., gemini-embedding-001). If not provided, defaults to gemini-embedding-001.
• input (string | array<string>): The text to embed.
• dimensions (number, optional): The output dimensionality.

Response Format:

{
  "object": "list",
  "data": [
    {
      "object": "embedding",
      "index": 0,
      "embedding": [0.1, 0.2, ...]
    }
  ],
  "model": "gemini-embedding-001"
}

Retrieve a list of available models. The response follows the OpenAI format where model names are stripped of the models/ prefix.

Response Format:

{
  "object": "list",
  "data": [
    {
      "id": "gemini-1.5-flash",
      "object": "model",
      "created": 0,
      "owned_by": ""
    }
  ]
}