waylaidwanderer/node-chatgpt-api

Support for the official ChatGPT model has been added! You can now use the gpt-3.5-turbo model with the official OpenAI API, using ChatGPTClient. This is the same model that ChatGPT uses, and it's the most powerful model available right now. Usage of this model is not free, however it is 10x cheaper (priced at $0.002 per 1k tokens) than text-davinci-003.

See OpenAI's post, Introducing ChatGPT and Whisper APIs for more information.

To use it, set modelOptions.model to gpt-3.5-turbo, and ChatGPTClient will handle the rest. The default model used in ChatGPTClient is now gpt-3.5-turbo. You can still set userLabel, chatGptLabel and promptPrefix (system instructions) as usual.

There may be a higher chance of your account being banned if you continue to automate chat.openai.com. Continue doing so at your own risk.

See demos/use-bing-client.js.

See demos/use-client.js.

See demos/use-browser-client.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 necessary when using `ChatGPTClient`, or `BingAIClient` in jailbreak mode.
    cacheOptions: {},
    // If set, `ChatGPTClient` and `BingAIClient` 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',
    chatGptClient: {
        // Your OpenAI API key (for `ChatGPTClient`)
        openaiApiKey: process.env.OPENAI_API_KEY || '',
        // (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.
            // The default model is `gpt-3.5-turbo`.
            model: 'gpt-3.5-turbo',
            // 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...".
        // (Optional) Set a custom name for the user
        // userLabel: 'User',
        // (Optional) Set a custom name for ChatGPT ("ChatGPT" by default)
        // chatGptLabel: 'Bob',
        // promptPrefix: 'You are Bob, a cowboy in Western times...',
        // A proxy string like "http://<ip>:<port>"
        proxy: '',
        // (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 conversation endpoint (private API server).
        // Warning: This will expose your access token to a third party. Consider the risks before using this.
        reverseProxyUrl: 'https://bypass.churchless.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: '',
        // A proxy string like "http://<ip>:<port>"
        proxy: '',
        // (Optional) Set to true to enable `console.debug()` logging
        debug: false,
    },
    // 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) Possible options: "chatgpt", "chatgpt-browser", "bing". (Default: "chatgpt")
        clientToUse: 'chatgpt',
        // (Optional) Generate titles for each conversation for clients that support it (only ChatGPTClient for now).
        // This will be returned as a `title` property in the first response of the conversation.
        generateTitles: false,
        // (Optional) Set this to allow changing the client or client options in POST /conversation.
        // To disable, set to `null`.
        perMessageClientOptionsWhitelist: {
            // The ability to switch clients using `clientOptions.clientToUse` will be disabled if `validClientsToUse` is not set.
            // To allow switching clients per message, you must set `validClientsToUse` to a non-empty array.
            validClientsToUse: ['bing', 'chatgpt', 'chatgpt-browser'], // values from possible `clientToUse` options above
            // The Object key, e.g. "chatgpt", is a value from `validClientsToUse`.
            // If not set, ALL options will be ALLOWED to be changed. For example, `bing` is not defined in `perMessageClientOptionsWhitelist` above,
            // so all options for `bingAiClient` will be allowed to be changed.
            // If set, ONLY the options listed here will be allowed to be changed.
            // In this example, each array element is a string representing a property in `chatGptClient` above.
            chatgpt: [
                'promptPrefix',
                'userLabel',
                'chatGptLabel',
                // Setting `modelOptions.temperature` here will allow changing ONLY the temperature.
                // Other options like `modelOptions.model` will not be allowed to be changed.
                // If you want to allow changing all `modelOptions`, define `modelOptions` here instead of `modelOptions.temperature`.
                'modelOptions.temperature',
            ],
        },
    },
    // Options for the CLI app
    cliOptions: {
        // (Optional) Possible options: "chatgpt", "bing".
        // clientToUse: 'bing',
    },
};

Start or continue a conversation. Optional parameters are only necessary for conversations that span multiple requests.

To configure which options can be changed per message (default: all), see the comments for perMessageClientOptionsWhitelist in settings.example.js. To allow changing clients, perMessageClientOptionsWhitelist.validClientsToUse must be set to a non-empty array as described in the example settings file.

To start a conversation with ChatGPT, send a POST request to the server's /conversation endpoint with a JSON body with parameters per Endpoints > POST /conversation above.

{
    "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": "an object containing the raw response from the client"
}

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."
}

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'; // use `@microsoft/fetch-event-source` instead if in a browser environment

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,
        // Any other parameters per `Endpoints > POST /conversation` above
    }),
};

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, see Method 1>', 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.
}

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

Instructions are provided below.