Updates

2023-02-19

I've added an experimental ChatGPTBrowserClient which depends on a reverse proxy server that makes use of a Cloudflare bypass, allowing you to talk to ChatGPT (chat.openai.com) without requiring browser automation. All you need is your access token from https://chat.openai.com/api/auth/session.

As always, please note that if you choose to go this route, you are exposing your access token to a closed-source third-party server. If you are concerned about this, you may choose to either use a free ChatGPT account to minimize risks, or continue using ChatGPTClient instead with the text-davinci-003 model.

Previous Updates

2023-02-15

The method we were using to access the ChatGPT raw models has been patched, unfortunately. Your options right now are to either use the official OpenAI API with the text-davinci-003 model (which costs money), or use a browser-based solution to interface with ChatGPT's backend (which is less powerful, more rate-limited and is not supported by this library at this time).

2023-02-11

With the help of @PawanOsman, we've figured out a way to continue using the ChatGPT raw models. To hopefully prevent losing access again, we've decided to provide reverse proxy servers compatible with the OpenAI API. I've updated ChatGPTClient to support using a reverse proxy server instead of the OpenAI API server. See Using a Reverse Proxy for more information on available proxy servers and how they work.

Please note that if you choose to go this route, you are exposing your access token to a closed-source third-party server. If you are concerned about this, you may choose to either use a free ChatGPT account to minimize risks, or continue using the official OpenAI API instead with the text-davinci-003 model.

2023-02-10

I've found a new working model for text-chat-davinci-002, text-chat-davinci-002-sh-alpha-aoruigiofdj83. This is the raw model that the ChatGPT Plus "Turbo" version uses. Responses are blazing fast. I've updated the library to use this model.

Bad timing; text-chat-davinci-002-sh-alpha-aoruigiofdj83 was removed shortly after, possibly due to a new model somewhere out there?

2023-02-09

Experience the power of Bing's GPT-4 version of ChatGPT with BingAIClient (experimental). The API server and CLI still need to be updated to support this, but you can use the client directly right now. Please note that if your account is still wait-listed, you will not be able to use this client.

2023-02-08

Even though text-chat-davinci-002-20221122 is back up again, it seems like it's constantly overloaded and returns a 429 error. It's likely that OpenAI only dedicated a small amount of resources to this model to prevent it being widely used by the public. Additionally, I've heard that newer versions are now access-locked to OpenAI employees and partners, so it's unlikely that we'll be able to find any workarounds until the model is officially released.

You may use the text-davinci-003 model instead as a drop-in replacement. Keep in mind that text-davinci-003 is not as good as text-chat-davinci-002 (which is trained via RHLF and fine-tuned to be a conversational AI), though results are still pretty good in most cases. Please note that using text-davinci-003 will cost you credits ($).

I will be re-adding support for the browser-based ChatGPT for the API server and CLI. Please star and watch this repository for updates.

2023-02-07

The roller coaster has reached the next stop. text-chat-davinci-002-20221122 is back up again.

Trying to use text-chat-davinci-002-20221122 with the OpenAI API now returns a 404 error. You may use the text-davinci-003 model instead as a drop-in replacement. Keep in mind that text-davinci-003 is not as good as text-chat-davinci-002 (which is trained via RHLF and fine-tuned to be a conversational AI), though results are still very good. Please note that using text-davinci-003 will cost you credits ($).

Please hold for further updates as we investigate further workarounds.

2023-02-02

Trying to use text-chat-davinci-002-20230126 with the OpenAI API now returns a 404 error. Someone has already found the new model name, but they are unwilling to share at this time. I will update this repository once I find the new model. If you have any leads, please open an issue or a pull request.

In the meantime, I've added support for models like text-davinci-003, which you can use as a drop-in replacement. Keep in mind that text-davinci-003 is not as good as text-chat-davinci-002 (which is trained via RHLF and fine-tuned to be a conversational AI), though results are still very good. Please note that using text-davinci-003 will cost you credits ($).

Discord user @pig#8932 has found a working text-chat-davinci-002 model, text-chat-davinci-002-20221122. I've updated the library to use this model.

ChatGPT API

A ChatGPT implementation with support for Bing's GPT-4 version of ChatGPT, plus the official ChatGPT model via OpenAI's API. Available as a Node.js module, REST API server, and CLI app.

This is an implementation of ChatGPT, with support for Bing's GPT-4 version of ChatGPT, plus the official ChatGPT raw model, text-chat-davinci-002.

About Bing's GPT-4 version of ChatGPT

An experimental client for Bing's GPT-4 version of ChatGPT is available in BingAIClient. It works much like ChatGPT, but it's powered by GPT-4 instead of GPT-3. For more information on its capabilities and limitations, see this Reddit comment.

About text-chat-davinci-002

The model name text-chat-davinci-002-20230126 was briefly leaked while I was inspecting the network requests made by the official ChatGPT website, and I discovered that it works with the OpenAI API. Since then, that model and others have been disabled, but I'm keeping this repo updated with the newer versions of text-chat-davinci-002 as we find them. Usage of this model currently does not cost any credits.

As far as I'm aware, I was the first one who discovered this, and usage of the model has since been implemented in libraries like acheong08/ChatGPT and transitive-bullshit/chatgpt-api as we collaborated and shared knowledge.

By itself, the model does not have any conversational support, so ChatGPTClient uses a cache to store conversations and pass them to the model as context. This allows you to have persistent conversations with ChatGPT in a nearly identical way to the official website.

Table of Contents

• Features
• Getting Started
  • Prerequisites
  • Usage
    • Module
    • API Server
    • CLI
  • Using a Reverse Proxy
• Caveats
• Contributing
• License

Features

• Experimental support for Bing's version of ChatGPT, powered by GPT-4.
• Support for the official ChatGPT raw model, text-chat-davinci-002, via OpenAI's API.
• Includes an API server (with Docker support) you can run to use ChatGPT in non-Node.js applications.
• Includes a ChatGPTClient and BingAIClient class that you can use in your own Node.js applications.
• Includes a CLI interface where you can chat with ChatGPT.
• (ChatGPTClient) Replicates chat threads from the official ChatGPT website (with conversation IDs and message IDs), with persistent conversations using Keyv.
  • Conversations are stored in memory by default, but you can optionally install a storage adapter to persist conversations to a database.
  • The keyv-file adapter is also included in this package, and can be used to store conversations in a JSON file if you're using the API server or CLI (see settings.example.js).
• (ChatGPTClient) Supports configurable prompt prefixes, and custom names for the user and ChatGPT.
  • In essence, this allows you to turn ChatGPT into a different character.
  • This is currently only configurable on a global level, but I plan to add support for per-conversation customization.

Getting Started

Prerequisites

• Node.js >= 16.0.0
• npm
• Docker (optional, for API server)
• OpenAI API key

Usage

Module

npm i @waylaidwanderer/chatgpt-api

BingAIClient

See demos/use-bing-client.js.

ChatGPTClient

See demos/use-client.js.

ChatGPTBrowserClient

See demos/use-browser-client.js.

API Server

Setup

You can install the package using

npm i -g @waylaidwanderer/chatgpt-api

then run it using chatgpt-api. This takes an optional --settings=<path_to_settings.js> parameter, or looks for settings.js in the current directory if not set, with the following contents:

settings.js

module.exports = {
    // Options for the Keyv cache, see https://www.npmjs.com/package/keyv.
    // This is used for storing conversations, and supports additional drivers (conversations are stored in memory by default).
    // Only applies when using `ChatGPTClient`.
    cacheOptions: {},
    // If set, `ChatGPTClient` will use `keyv-file` to store conversations to this JSON file instead of in memory.
    // However, `cacheOptions.store` will override this if set
    storageFilePath: process.env.STORAGE_FILE_PATH || './cache.json',
    // Your OpenAI API key (for `ChatGPTClient`)
    openaiApiKey: process.env.OPENAI_API_KEY || '',
    chatGptClient: {
        // (Optional) Support for a reverse proxy for the completions endpoint (private API server).
        // Warning: This will expose your `openaiApiKey` to a third party. Consider the risks before using this.
        // reverseProxyUrl: 'https://chatgpt.hato.ai/completions',
        // (Optional) Parameters as described in https://platform.openai.com/docs/api-reference/completions
        modelOptions: {
            // You can override the model name and any other parameters here.
            // model: 'text-chat-davinci-002-20221122',
            // Set max_tokens here to override the default max_tokens of 1000 for the completion.
            // max_tokens: 1000,
        },
        // (Optional) Davinci models have a max context length of 4097 tokens, but you may need to change this for other models.
        // maxContextTokens: 4097,
        // (Optional) You might want to lower this to save money if using a paid model like `text-davinci-003`.
        // Earlier messages will be dropped until the prompt is within the limit.
        // maxPromptTokens: 3097,
        // (Optional) Set custom instructions instead of "You are ChatGPT...".
        // promptPrefix: 'You are Bob, a cowboy in Western times...',
        // (Optional) Set a custom name for the user
        // userLabel: 'User',
        // (Optional) Set a custom name for ChatGPT
        // chatGptLabel: 'ChatGPT',
        // (Optional) Set to true to enable `console.debug()` logging
        debug: false,
    },
    // Options for the Bing client
    bingAiClient: {
        // Necessary for some people in different countries, e.g. China (https://cn.bing.com)
        host: '',
        // The "_U" cookie value from bing.com
        userToken: '',
        // If the above doesn't work, provide all your cookies as a string instead
        cookies: '',
        // A proxy string like "http://<ip>:<port>"
        proxy: '',
        // (Optional) Set to true to enable `console.debug()` logging
        debug: false,
    },
    chatGptBrowserClient: {
        // (Optional) Support for a reverse proxy for the completions endpoint (private API server).
        // Warning: This will expose your access token to a third party. Consider the risks before using this.
        reverseProxyUrl: 'https://chatgpt.duti.tech/api/conversation',
        // Access token from https://chat.openai.com/api/auth/session
        accessToken: '',
        // Cookies from chat.openai.com (likely not required if using reverse proxy server).
        cookies: '',
        // (Optional) Set to true to enable `console.debug()` logging
        // debug: true,
    },
    // Options for the API server
    apiOptions: {
        port: process.env.API_PORT || 3000,
        host: process.env.API_HOST || 'localhost',
        // (Optional) Set to true to enable `console.debug()` logging
        debug: false,
        // (Optional) Set to "bing" to use `BingAIClient` instead of `ChatGPTClient`.
        // clientToUse: 'bing',
    },
    // Options for the CLI app
    cliOptions: {
        // (Optional) Set to "bing" to use `BingAIClient` instead of `ChatGPTClient`.
        // clientToUse: 'bing',
    },
};

Alternatively, you can install and run the package directly.

1. Clone this repository: git clone https://github.com/waylaidwanderer/node-chatgpt-api
2. Install dependencies with npm install (if not using Docker)
3. Rename settings.example.js to settings.js in the root directory and change the settings where required.
4. Start the server:
   • using npm start or npm run server (if not using Docker)
   • using docker-compose up (requires Docker)

Usage

Method 1 (POST)

To start a conversation with ChatGPT, send a POST request to the server's /conversation endpoint with a JSON body in the following format. Optional parameters are only necessary for conversations that span multiple requests:

{
    "message": "Hello, how are you today?",
    "conversationId": "your-conversation-id (optional)",
    "parentMessageId": "your-parent-message-id (optional, for `ChatGPTClient` only)",
    "conversationSignature": "your-conversation-signature (optional, for `BingAIClient` only)",
    "clientId": "your-client-id (optional, for `BingAIClient` only)",
    "invocationId": "your-invocation-id (optional, for `BingAIClient` only)",
}

The server will return a JSON object containing ChatGPT's response:

// HTTP/1.1 200 OK
{
    "response": "I'm doing well, thank you! How are you?",
    "conversationId": "your-conversation-id",
    "messageId": "response-message-id (for `ChatGPTClient` only)",
    "conversationSignature": "your-conversation-signature (for `BingAIClient` only)",
    "clientId": "your-client-id (for `BingAIClient` only)",
    "invocationId": "your-invocation-id (for `BingAIClient` only - pass this new value back into subsequent requests as-is)",
    "details": "additional details about the AI's response (for `BingAIClient` only)"
}

If the request is unsuccessful, the server will return a JSON object with an error message.

If the request object is missing a required property (e.g. message):

// HTTP/1.1 400 Bad Request
{
    "error": "The message parameter is required."
}

If there was an error sending the message to ChatGPT:

// HTTP/1.1 503 Service Unavailable
{
    "error": "There was an error communicating with ChatGPT."
}

Method 2 (SSE)

You can set "stream": true in the request body to receive a stream of tokens as they are generated.

import { fetchEventSource } from '@waylaidwanderer/fetch-event-source';

const opts = {
    method: 'POST',
    headers: {
        'Content-Type': 'application/json',
    },
    body: JSON.stringify({
        "message": "Write a poem about cats.",
        "conversationId": "your-conversation-id (optional)",
        "parentMessageId": "your-parent-message-id (optional)",
        "stream": true
    }),
};

See demos/use-api-server-streaming.js for an example of how to receive the response as it's generated. You will receive one token at a time, so you will need to concatenate them yourself.

Successful output:

{ data: '', event: '', id: '', retry: 3000 }
{ data: 'Hello', event: '', id: '', retry: undefined }
{ data: '!', event: '', id: '', retry: undefined }
{ data: ' How', event: '', id: '', retry: undefined }
{ data: ' can', event: '', id: '', retry: undefined }
{ data: ' I', event: '', id: '', retry: undefined }
{ data: ' help', event: '', id: '', retry: undefined }
{ data: ' you', event: '', id: '', retry: undefined }
{ data: ' today', event: '', id: '', retry: undefined }
{ data: '?', event: '', id: '', retry: undefined }
{ data: '<result JSON here>', event: 'result', id: '', retry: undefined }
{ data: '[DONE]', event: '', id: '', retry: undefined }
// Hello! How can I help you today?

Error output:

const message = {
  data: '{"code":503,"error":"There was an error communicating with ChatGPT."}',
  event: 'error',
  id: '',
  retry: undefined
};

if (message.event === 'error') {
  console.error(JSON.parse(message.data).error); // There was an error communicating with ChatGPT.
}

Notes

• Method 1 is simple, but Time to First Byte (TTFB) is long.
• Method 2 uses a non-standard implementation of server-sent event API; you should import fetch-event-source first and use POST method.

CLI

Setup

Follow the same setup instructions for the API server, creating settings.js.

Usage

If installed globally:

chatgpt-cli

If installed locally:

npm run cli

ChatGPT's responses are automatically copied to your clipboard, so you can paste them into other applications.

Using a Reverse Proxy

As shown in the examples above, you can set reverseProxyUrl in ChatGPTClient's options to use a reverse proxy server instead of the official ChatGPT API. For now, this is the only way to use the ChatGPT raw models. This method has been patched, but you may still want to use a reverse proxy for other reasons.

Instructions

How does it work? Simple answer: ChatGPTClient > reverse proxy > OpenAI server. The reverse proxy server does some magic under the hood to access the raw model directly via OpenAI's server and then returns the response to ChatGPTClient.

Instructions are provided below.

https://chatgpt.hato.ai/completions (mine, currently offline)

Instructions

1. Get your ChatGPT access token from https://chat.openai.com/api/auth/session (look for the accessToken property).
   • This is NOT the same thing as the session token.
   • Automatically fetching or refreshing your ChatGPT access token is not currently supported by this library. Please handle this yourself for now.
2. Set reverseProxyUrl to https://chatgpt.hato.ai/completions in settings.js > chatGptClient or ChatGPTClient's options.
3. Set the "OpenAI API key" parameter (e.g. settings.openaiApiKey) to the ChatGPT access token you got in step 1.
4. Set the model to text-davinci-002-render, text-davinci-002-render-paid, or text-davinci-002-render-sha depending on which ChatGPT models that your account has access to. Models must be a ChatGPT model name, not the raw model name, and you cannot use a model that your account does not have access to.
   • You can check which ones you have access to by opening DevTools and going to the Network tab. Refresh the page and look at the response body for https://chat.openai.com/backend-api/models.

Notes

• Since this is my server, I can guarantee that no logging or tracking is done. I can see general usage stats, but I cannot see any of your completions. Whether you trust me on this or not is up to you.
• Non-streaming responses over 60s are not supported. Use stream: true (API) or onProgress (client) as a workaround.
• Rate limit of 10 requests per second.

https://chatgpt.pawan.krd/api/completions (@PawanOsmon, currently offline)

Instructions

1. Get your ChatGPT access token from https://chat.openai.com/api/auth/session (look for the accessToken property).
   • This is NOT the same thing as the session token.
   • Automatically fetching or refreshing your ChatGPT access token is not currently supported by this library. Please handle this yourself for now.
2. Set reverseProxyUrl to https://chatgpt.pawan.krd/api/completions in settings.js > chatGptClient or ChatGPTClient's options.
3. Set the "OpenAI API key" parameter (e.g. settings.openaiApiKey) to the ChatGPT access token you got in step 1.
4. Set the model to text-davinci-002-render, text-davinci-002-render-paid, or text-davinci-002-render-sha depending on which ChatGPT models that your account has access to. Models must be a ChatGPT model name, not the raw model name, and you cannot use a model that your account does not have access to.
   • You can check which ones you have access to by opening DevTools and going to the Network tab. Refresh the page and look at the response body for https://chat.openai.com/backend-api/models.

Notes

• Non-streaming responses over 60s are not supported. Use stream: true (API) or onProgress (client) as a workaround.
• Rate limit of 50 requests per 15 seconds.

Caveats

Since text-chat-davinci-002 is ChatGPT's raw model, I had to do my best to replicate the way the official ChatGPT website uses it. After extensive testing and comparing responses, I believe that the model used by ChatGPT has some additional fine-tuning. This means my implementation or the raw model may not behave exactly the same in some ways:

• Conversations are not tied to any user IDs, so if that's important to you, you should implement your own user ID system.

• ChatGPT's model parameters (temperature, frequency penalty, etc.) are unknown, so I set some defaults that I thought would be reasonable.

• Conversations are limited to roughly the last 3000 tokens, so earlier messages may be forgotten during longer conversations.

  • This works in a similar way to ChatGPT, except I'm pretty sure they have some additional way of retrieving context from earlier messages when needed (which can probably be achieved with embeddings, but I consider that out-of-scope for now).

• It is well known that, as part of the fine-tuning, ChatGPT had the following preamble:

  "You are ChatGPT, a large language model trained by OpenAI. You answer as concisely as possible for each response (e.g. don’t be verbose). It is very important that you answer as concisely as possible, so please remember this. If you are generating a list, do not have too many items. Keep the number of items short. Knowledge cutoff: 2021-09 Current date: 2023-01-31"

  As OpenAI updates ChatGPT, this preamble may also change. The default prompt prefix in my implementation attempts to replicate a similar behavior to the current ChatGPT model.

Contributing

If you'd like to contribute to this project, please create a pull request with a detailed description of your changes.

License

This project is licensed under the MIT License.