GeminiAI Compatible with OpenAI
GeminiAI is a developer-friendly large language model with a beautiful interface, powerful features, and a generous daily free quota that meets everyday usage needs.
However, it has some inconveniences, such as requiring a proxy to access and having an API that is not compatible with the OpenAI SDK.
To solve these issues and achieve compatibility with OpenAI, I wrote a piece of JavaScript code and deployed it on Cloudflare, binding it to my own domain. This allows using Gemini in China without a proxy while maintaining compatibility with OpenAI. In any tool that uses OpenAI, simply replace the API endpoint and API key (SK).
Create a Worker on Cloudflare
If you don't have a Cloudflare account yet, register one first (it's free). Registration link: https://dash.cloudflare.com/ After logging in, remember to bind your own domain; otherwise, you won't be able to access it without a proxy.
After logging in, find Compute (Workers) in the left sidebar and click it, then click the Create button.
On the page that appears, click Create Worker.
Then click Deploy in the bottom right corner to complete the Worker creation.
Edit the Code
The code below is key to achieving OpenAI compatibility. Please copy it and replace the default generated code in the Worker.
On the page after deployment, click Edit Code.
Delete all the code on the left, then copy and paste the code below, and finally click Deploy in the top right corner.
Copy the code below
export default {
async fetch (request) {
if (request.method === "OPTIONS") {
return handleOPTIONS();
}
const errHandler = (err) => {
console.error(err);
return new Response(err.message, fixCors({ status: err.status ?? 500 }));
};
try {
const auth = request.headers.get("Authorization");
const apiKey = auth?.split(" ")[1];
const assert = (success) => {
if (!success) {
throw new HttpError("The specified HTTP method is not allowed for the requested resource", 400);
}
};
const { pathname } = new URL(request.url);
if(!pathname.endsWith("/chat/completions")){
return new Response("hello")
}
assert(request.method === "POST");
return handleCompletions(await request.json(), apiKey).catch(errHandler);
} catch (err) {
return errHandler(err);
}
}
};
class HttpError extends Error {
constructor(message, status) {
super(message);
this.name = this.constructor.name;
this.status = status;
}
}
const fixCors = ({ headers, status, statusText }) => {
headers = new Headers(headers);
headers.set("Access-Control-Allow-Origin", "*");
return { headers, status, statusText };
};
const handleOPTIONS = async () => {
return new Response(null, {
headers: {
"Access-Control-Allow-Origin": "*",
"Access-Control-Allow-Methods": "*",
"Access-Control-Allow-Headers": "*",
}
});
};
const BASE_URL = "https://generativelanguage.googleapis.com";
const API_VERSION = "v1beta";
// https://github.com/google-gemini/generative-ai-js/blob/cf223ff4a1ee5a2d944c53cddb8976136382bee6/src/requests/request.ts#L71
const API_CLIENT = "genai-js/0.21.0"; // npm view @google/generative-ai version
const makeHeaders = (apiKey, more) => ({
"x-goog-api-client": API_CLIENT,
...(apiKey && { "x-goog-api-key": apiKey }),
...more
});
const DEFAULT_MODEL = "gemini-2.0-flash-exp";
async function handleCompletions (req, apiKey) {
let model = DEFAULT_MODEL;
if(req.model.startsWith("gemini-")) {
model = req.model;
}
const TASK = "generateContent";
let url = `${BASE_URL}/${API_VERSION}/models/${model}:${TASK}`;
const response = await fetch(url, {
method: "POST",
headers: makeHeaders(apiKey, { "Content-Type": "application/json" }),
body: JSON.stringify(await transformRequest(req)), // try
});
let body = response.body;
if (response.ok) {
let id = generateChatcmplId();
body = await response.text();
body = processCompletionsResponse(JSON.parse(body), model, id);
}
return new Response(body, fixCors(response));
}
const harmCategory = [
"HARM_CATEGORY_HATE_SPEECH",
"HARM_CATEGORY_SEXUALLY_EXPLICIT",
"HARM_CATEGORY_DANGEROUS_CONTENT",
"HARM_CATEGORY_HARASSMENT",
"HARM_CATEGORY_CIVIC_INTEGRITY",
];
const safetySettings = harmCategory.map(category => ({
category,
threshold: "BLOCK_NONE",
}));
const fieldsMap = {
stop: "stopSequences",
n: "candidateCount",
max_tokens: "maxOutputTokens",
max_completion_tokens: "maxOutputTokens",
temperature: "temperature",
top_p: "topP",
top_k: "topK",
frequency_penalty: "frequencyPenalty",
presence_penalty: "presencePenalty",
};
const transformConfig = (req) => {
let cfg = {};
for (let key in req) {
const matchedKey = fieldsMap[key];
if (matchedKey) {
cfg[matchedKey] = req[key];
}
}
cfg.responseMimeType = "text/plain";
return cfg;
};
const transformMsg = async ({ role, content }) => {
const parts = [];
if (!Array.isArray(content)) {
parts.push({ text: content });
return { role, parts };
}
for (const item of content) {
switch (item.type) {
case "text":
parts.push({ text: item.text });
break;
case "input_audio":
parts.push({
inlineData: {
mimeType: "audio/" + item.input_audio.format,
data: item.input_audio.data,
}
});
break;
default:
throw new TypeError(`Unknown "content" item type: "${item.type}"`);
}
}
if (content.every(item => item.type === "image_url")) {
parts.push({ text: "" });
}
return { role, parts };
};
const transformMessages = async (messages) => {
if (!messages) { return; }
const contents = [];
let system_instruction;
for (const item of messages) {
if (item.role === "system") {
delete item.role;
system_instruction = await transformMsg(item);
} else {
item.role = item.role === "assistant" ? "model" : "user";
contents.push(await transformMsg(item));
}
}
if (system_instruction && contents.length === 0) {
contents.push({ role: "model", parts: { text: " " } });
}
return { system_instruction, contents };
};
const transformRequest = async (req) => ({
...await transformMessages(req.messages),
safetySettings,
generationConfig: transformConfig(req),
});
const generateChatcmplId = () => {
const characters = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789";
const randomChar = () => characters[Math.floor(Math.random() * characters.length)];
return "chatcmpl-" + Array.from({ length: 29 }, randomChar).join("");
};
const reasonsMap = {
"STOP": "stop",
"MAX_TOKENS": "length",
"SAFETY": "content_filter",
"RECITATION": "content_filter"
};
const SEP = "\n\n|>";
const transformCandidates = (key, cand) => ({
index: cand.index || 0,
[key]: {
role: "assistant",
content: cand.content?.parts.map(p => p.text).join(SEP) },
logprobs: null,
finish_reason: reasonsMap[cand.finishReason] || cand.finishReason,
});
const transformCandidatesMessage = transformCandidates.bind(null, "message");
const transformCandidatesDelta = transformCandidates.bind(null, "delta");
const transformUsage = (data) => ({
completion_tokens: data.candidatesTokenCount,
prompt_tokens: data.promptTokenCount,
total_tokens: data.totalTokenCount
});
const processCompletionsResponse = (data, model, id) => {
return JSON.stringify({
id,
choices: data.candidates.map(transformCandidatesMessage),
created: Math.floor(Date.now()/1000),
model,
object: "chat.completion",
usage: transformUsage(data.usageMetadata),
});
};Bind a Domain
After deployment, Cloudflare provides a secondary subdomain, but it cannot be accessed normally in China. Therefore, you need to bind your own domain to enable proxy-free access.
After deployment, click Back on the left.
Then find Settings -> Domains and Routes, and click Add.
As shown below, add the domain you have already hosted on Cloudflare.
After completion, you can use this domain to access Gemini.
Access Gemini Using the OpenAI SDK
from openai import OpenAI, APIConnectionError
model = OpenAI(api_key='Your Gemini API Key', base_url='https://your-custom-domain.com')
response = model.chat.completions.create(
model='gemini-2.0-flash-exp',
messages=[
{
'role': 'user',
'content': 'Who are you?'},
]
)
print(response.choices[0])Returns:
Choice(finish_reason='stop', index=0, logprobs=None, message=ChatCompletionMessage(content='I am a large language model trained by Google.\n', refusal=None, role='assistant', audio=None, function_call=None, tool_calls=None))Use in Other OpenAI-Compatible Tools
Find the location in the tool where OpenAI information is configured. Change the API endpoint to your custom domain added in Cloudflare, replace the SK with your Gemini API Key, and set the model to gemini-2.0-flash-exp.
Direct Access Using Requests
If you don't use the OpenAI SDK, you can also access it directly using the requests library.
import requests
payload={
"model":"gemini-1.5-flash",
"messages":[{
"role":"user",
"content":[{"type":"text","text":"Who are you?"}]
}]
}
res=requests.post('https://xxxx.com/chat/completions',headers={"Authorization":"Bearer Your Gemini API Key","Content-Type":"application:/json"},json=payload)
print(res.json())Output:
Related Resources
- Source code modified from the project PublicAffairs/openai-gemini
- GeminiAI Documentation