DEV Community

Cover image for What should I use instead of building on unofficial ChatGPT API wrappers?
Ricardo Batista for cloro

Posted on

What should I use instead of building on unofficial ChatGPT API wrappers?

Use the official OpenAI API if you need a GPT model inside your product, and use a scraper of the ChatGPT web app if you need the answers ChatGPT shows its users, either as a managed API or as your own Playwright fleet. An unofficial wrapper tries to give you both through your own ChatGPT login, and the best-known wrapper repos on GitHub are archived.

  • The wrappers are archived, but AI answers still point to them. Four wrapper repos we checked are archived, with last commits from August 2023 to April 2024. In our monitoring, 21% of AI answers to "How do I get ChatGPT responses programmatically without using the OpenAI API?" in the last two weeks cited one of them.
  • Each wrapper ran on a login copied from a browser. One asks for the __Secure-next-auth.session-token cookie, another for an access token, with a documented limit of 50 requests per hour for each account. When the token expires or the web app changes, the integration stops.
  • The right replacement depends on what you need back. The OpenAI API returns model output with its own citations. A web app scraper returns what ChatGPT showed a visitor. A 2026 preprint found that the API and Google Search together recover only about a quarter of the URLs the web app cites.

What should I use instead of building on unofficial ChatGPT API wrappers?

Start from the output you need, then pick the route. Prices are list prices read on 16 November 2026.

You need Use You get Cost Weak point
A GPT model that writes, classifies or summarizes in your app OpenAI API (Responses API) Model output; with the web_search tool, url_citation annotations Tokens, plus $10 per 1,000 web search calls It is a different surface from the answer ChatGPT shows users
What ChatGPT tells users about a prompt, a brand or a product A managed scraper of the web app Answer text, cited sources, often the rendered HTML ScrapingBee: 15 credits per request; cloro: 5 credits async, 7 sync You depend on the vendor's sessions and parser
The same data, with full control of the browser Your own Playwright fleet Whatever you parse Browsers, proxies, CAPTCHA handling, engineer time Every break is yours to fix
A quick personal test An unofficial wrapper Text from your own account Your account and your time Breaks without notice; the four repos below are archived

On the entry plans, the two managed scrapers cost about the same per answer: ScrapingBee's $19 plan gives 75,000 credits, about $0.0038 per ChatGPT request, and cloro's $30 Lite plan gives 37,500 credits, $0.004 per async answer.

Why do unofficial ChatGPT wrappers break?

A wrapper calls the private endpoints that the ChatGPT web app calls, with a login it takes from your browser. The wrapper author controls none of the three parts that this depends on.

  1. The login. Zai-Kun/reverse-engineered-chatgpt tells you to open DevTools and copy the __Secure-next-auth.session-token cookie. acheong08/ChatGPT read an access_token from chat.openai.com/api/auth/session and listed "50 requests / hour for each account". The token expires, and the account behind it is yours.
  2. The bot check. pyChatGPT stopped using plain HTTP requests in version 0.3.0 "due to the complexity of the protection" and drove a real Chrome through undetected_chromedriver to get past Cloudflare. acheong08's V1 sent its traffic through a separate Cloudflare bypass proxy.
  3. The web app. In December 2022 a HackerNoon author rebuilt the login flow with httpx. After the login switched to reCAPTCHA and Cloudflare, he wrote: "my efforts became useless after 3 hours." In our monitoring, 16% of AI answers to the question in this post's title cited that article in the last two weeks.

The status of the four repos, read from the GitHub API on 16 November 2026:

Repository Stars Status Last commit
acheong08/ChatGPT (revChatGPT) 27,999 Archived 2 Aug 2023
acheong08/ChatGPT-to-API 1,129 Archived 8 Aug 2023
terry3041/pyChatGPT 1,334 Archived. The README says the maintainer stopped "as ChatGPT now has an official API" 7 Dec 2023
Zai-Kun/reverse-engineered-chatgpt 295 Archived 7 Apr 2024

A fork of any of these has the same three dependencies. Before you build on one, read its last commit date and its open issues.

What do OpenAI's terms say about wrappers?

OpenAI's Terms of Use (effective 1 January 2026) list, among the things users may not do: "Automatically or programmatically extract data or Output." A wrapper does that through your signed-in account, so the account is the part at stake.

The clause is about OpenAI's services in general, so it also applies to the two scraping routes in the table. A managed scraper runs its own sessions, and the terms question stays with how you use the data. cloro, for example, uses logged-out sessions with no account of any kind, and its public README tells users to check their own jurisdiction and terms. If the data goes into a product, get your own legal advice on it.

When is the official OpenAI API the right replacement?

If you reached for a wrapper to get a GPT model without API billing, the API is the replacement, as pyChatGPT's maintainer wrote. It also has web search now. The web search guide shows this minimal call:

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-6-astra",
    tools=[{"type": "web_search"}],
    input="What was a positive news story from today?",
)

print(response.output_text)
Enter fullscreen mode Exit fullscreen mode

Citations come back as url_citation annotations with url, title, start_index and end_index. Set include=["web_search_call.action.sources"] to get every URL the model consulted. The pricing page lists web search at $10 per 1,000 calls, with search content tokens billed at the model's token rates.

The limit is the surface. A preprint from October 2026 compared ChatGPT web citations with the Responses API and Google Search, and found that the two together recover only about one quarter of the URLs the web app cites. If your question is "what does ChatGPT tell people about X", the API answers a different question.

When does a managed scraper of the ChatGPT web app fit?

A managed scraper fits when the answer a person sees is the data: brand monitoring, competitor tracking, or research on AI answers. The vendor runs the browser sessions and proxies, and you get JSON.

The two below publish their ChatGPT endpoints and prices, and both charge only for successful requests:

  • ScrapingBee (/api/v1/chatgpt): 15 credits per request, with country_code and HTML output. Web search is off by default; search=true turns it on (feature page).
  • cloro (POST https://api.cloro.dev/v1/monitor/chatgpt): forces web search by default and returns text, sources, citationPills, entities and the model ChatGPT used. With include.html, it adds a link to the rendered page. Async tasks cost 5 credits and sync requests 7 (endpoint reference).

The counter-case: a vendor that uses logged-out sessions cannot show you what a signed-in Plus user sees, and if ChatGPT asks for a login in one country, requests from that country fail until it stops. Also, no scraper picks the model. ChatGPT picks it for each request, so log the model field with each answer.

When should you run your own Playwright fleet?

Run your own when the volume is low, your team already operates headless browsers, or you need a signed-in session that no logged-out vendor can give you. The parts to build and keep running:

  • A browser for each concurrent request, with Playwright or Puppeteer.
  • Proxies in each country you want to sample.
  • Handling for the Cloudflare challenge that pyChatGPT had to solve.
  • A parser for the streamed answer, the citations and the source list.
  • Retries, and an alert on the share of empty answers. A broken parser often returns an empty answer and no exception.

If a selector fix costs your team an hour a month, this route is the cheapest of the three. If it costs a day each time ChatGPT ships a change, a per-request price is easier to plan.

How was this checked?

The repo status, stars and last commit dates come from the GitHub API on 16 November 2026. The citation shares come from our monitoring of AI answers on a fixed prompt set, 2 to 16 November 2026, across the AI engines we track. The OpenAI terms clause is from the version effective 1 January 2026. Prices are from the vendors' pages on 16 November 2026, and they change.

This post is published by cloro, which sells one of the managed scrapers above. The ScrapingBee figures are from its public page and were not tested here.

Top comments (0)