Internet ChatGPT Plugins
Last reviewed
May 9, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 2,498 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
May 9, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 2,498 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: ChatGPT Plugins, ChatGPT Plugin Categories and Internet
Internet ChatGPT Plugins were a loose grouping of third-party and first-party extensions inside the ChatGPT plugin store that gave the assistant some way to reach the live web. The grouping covered four use cases: general web browsing, fetching content from a specific URL, semantic or keyword search across the open internet, and scraping structured data from web pages. Plugins in this category sat alongside others such as search engine ChatGPT plugins and web development ChatGPT plugins, but they were defined less by a formal store taxonomy and more by what users wanted from them: a way for a model trained on a static corpus to read pages it had never seen.
The plugin platform launched in alpha on March 23, 2023, with broader rollout to ChatGPT Plus subscribers beginning around May 12, 2023. OpenAI wound the platform down on March 19, 2024, when no new conversations could start with plugins, and shut it off on April 9, 2024. Plugins were superseded by Custom GPTs, launched November 6, 2023, and the GPT Store, opened January 10, 2024. Live web access was eventually folded into ChatGPT itself as a built-in feature.
The base ChatGPT model in early 2023 had a knowledge cutoff in 2021, with later GPT-4 builds extending the cutoff partway into 2023. Anything more recent, including news, prices, sports scores, freshly published research, and changing reference pages, was outside what the model could answer from memory. Internet plugins existed to close that gap. They let the assistant make outbound HTTP requests, read what came back, and fold the result into a normal chat reply.
A second motivation was citation. The plain model could state facts but had no native way to point at the source it was relying on. Plugins that fetched a real URL could pass that URL back into the conversation, so users could click through and verify. The first-party browsing plugin returned source links with its answers, and several third-party plugins followed the same pattern.
A third motivation was working with documents the user already had in mind. An internet plugin could pull a specific blog post, hosted PDF, or known URL into the model's context window without the user copying and pasting the text. This workflow fed into later features like ChatGPT's native PDF attachments and direct URL handling.
Every ChatGPT plugin was a thin convention over a standard web API. A developer hosted a JSON manifest at /.well-known/ai-plugin.json on their domain, paired with an OpenAPI specification describing the available endpoints. ChatGPT downloaded both files, summarized them into a system prompt, and let the model decide when to call the API based on the user's request.
For an internet plugin, the API was usually a wrapper around either an HTTP fetcher (curl, headless browser, scraping service) or a search backend (Bing, Google via SERP API, a custom neural index). When the user asked a question that needed live information, the model wrote a structured call, the plugin server executed it, and the response, often trimmed to fit token budgets, came back as tool output. The model then wrote the user-facing reply on top of that text.
OpenAI's browsing plugin added constraints on top of the generic pattern. It used the Bing Search API as its index, only issued GET requests, and avoided form submissions or any action that could change state on the target site. That made it closer to a read-only crawler than a general-purpose web agent.
The most significant entry in the internet category was OpenAI's own plugin, sometimes labeled "Browsing" and later marketed as "Browse with Bing." It shipped on the first day of the plugin alpha, March 23, 2023, alongside a separate code interpreter and a small group of partner plugins. It used Bing as the underlying search and retrieval layer, which gave it global coverage without OpenAI having to operate its own crawler.
The browsing plugin had a turbulent history. OpenAI temporarily disabled it on July 3, 2023, after users demonstrated it could fetch text from articles that publishers had placed behind paywalls. The feature returned on September 27, 2023, with new safeguards: it now respected robots.txt directives, identified itself with a documented user-agent string, and gave site operators a way to block or rate-limit ChatGPT's traffic. The relaunch was officially branded "Browse with Bing" and made available to ChatGPT Plus and Enterprise users.
After the plugin platform was retired in 2024, browsing did not disappear. Instead, OpenAI rebuilt the same capability as a native ChatGPT feature, eventually folding it into the default model behavior so the assistant could choose to look something up without the user having to enable a separate tool.
Beyond OpenAI's own browser, the internet category was populated by a mix of independent developer plugins. The grouping below is descriptive rather than official; the plugin store's own categories shifted over time and many plugins straddled more than one.
These plugins took a URL or a set of URLs from the user, fetched the content, and returned a parsed version that ChatGPT could read.
| Plugin | Function |
|---|---|
| WebPilot | Fetched a single URL or a list of URLs, returned cleaned text, supported summarization and translation of fetched pages. Submitted in April 2023, briefly removed for review, relaunched July 18, 2023. |
| BrowserOp (formerly BrowserPilot) | Fetched specific webpages or searched the web on the user's behalf, marketed for current-events questions. |
| Link Reader | Read a wide range of link types including ordinary webpages, PDFs, and other document formats. |
| Access Link | Retrieved information from any web link supplied in the chat. |
| Web Requests | Acted as a generic HTTP client driven by the model, framed as "your assistant's web browser." |
| Chat With Many URLs | Accepted multiple URLs in one message and let the user query their combined text content. |
| ChatWithWebsite | Loaded a target site and let the user hold a conversation grounded in that site's pages. |
These plugins overlapped heavily with each other and with the first-party browser. Their differentiators were usually around how they parsed pages, how they handled long documents, and how aggressively they followed links from the original URL.
A second cluster wrapped a search engine rather than a single URL. The user typed a natural-language query, the plugin issued a search behind the scenes, and ChatGPT received a list of candidate pages plus snippets to read.
| Plugin | Search backend or focus |
|---|---|
| KeyMate.AI Search | Custom search service that combined keyword and AI-driven retrieval, often listed alongside the official browsing plugin and WebPilot as one of the three common ways to get current information. |
| Metaphor (later Exa) | Neural search engine designed to retrieve high quality links by meaning rather than keyword, useful for content discovery and recommendations. |
| AI Browse and Search | Combined a Google-style search step with page browsing in a single plugin flow. |
| Browser / BrowserPilot | Bundled real-time search results with content extraction across multiple URLs. |
| VoxScript | Searched YouTube transcripts, financial data, and Google search results in one plugin, popular for current-events grounding. |
The search-oriented plugins were how many ChatGPT Plus users got around the rolling outages of "Browse with Bing" during mid-2023. When the official browser was disabled, KeyMate, WebPilot, and Metaphor stayed available and absorbed much of the demand.
A smaller cluster of internet plugins specialized in pulling structured data out of a webpage rather than reading it as prose.
| Plugin | Specialty |
|---|---|
| Apify | Scraped Google search results, crawled websites, and queried discovered pages, exposing Apify's existing scraping infrastructure to the chat. |
| ScrapingBee | Combined a Google search step with a full-page scraper, returning rendered content for sites that needed JavaScript execution. |
| Aaron Browser | Scraped data from multiple website URLs, designed for crawling and content aggregation tasks. |
| Scrapee | Extracted text content from arbitrary webpages. |
| Scraper | A simpler URL-in, content-out scraper aimed at quick lookups. |
Scraping plugins blurred the line between the internet category and the web development ChatGPT plugins category, since they were often used by developers building data pipelines on top of ChatGPT rather than by general users asking factual questions.
Several plugins were not about generic browsing but still relied on fetching content from the open internet. They appeared in or adjacent to the internet grouping in many third-party plugin lists.
| Plugin | Internet-facing role |
|---|---|
| AskYourPDF | Accepted a URL pointing at a PDF, downloaded the file, indexed it, and let the user ask questions about its contents. |
| ChatWithPDF | Similar to AskYourPDF, query and analyze PDF documents linked from the web or from cloud drives. |
| ChatWithVideo | Took a YouTube video URL, parsed the transcript, and let the user ask questions about the video. |
| ScholarAI | Searched scientific literature databases for peer-reviewed papers and returned abstracts and links. |
| World News | Pulled news headlines from multiple online sources for current-events summaries. |
These plugins sat at the intersection of internet access and a more specific use case: research, video, news, or document Q and A. They mattered for the internet category mainly because they each made one specific kind of online content reachable from inside the chat.
The internet category overlapped with several others in the plugin store, and the boundaries were never strict. Search engine ChatGPT plugins covered tools whose primary job was to query a search index; Metaphor and KeyMate fit more naturally there, while WebPilot and Link Reader fit under generic internet access. Web development ChatGPT plugins covered tools for building, deploying, and inspecting websites, and scraping plugins like Apify and ScrapingBee straddled both categories. Research and data plugins like ScholarAI used the internet as a backbone but were classified by the document type they handled. The ChatGPT plugin categories overview tracks the broader taxonomy and how it shifted as the store grew from around 70 plugins at the May 2023 broad launch to several hundred by the time the platform was wound down.
Internet plugins had a few persistent issues during their lifetime.
Context budgets were tight. The plugin platform ran on GPT-4 with an 8,000 or 32,000 token window depending on the user's tier, and a single fetched webpage could exceed those limits. Plugins handled overflow differently: some truncated, some summarized, some returned only the visible text and dropped sidebars and navigation. A question that worked against one plugin would sometimes fail against another simply because of how the page had been chunked.
Paywalls and licensed content were a recurring policy problem. The first-party browsing plugin was disabled for almost three months in 2023 because users demonstrated it could fetch text from publisher sites that should have been gated. Third-party plugins that scraped pages without enforcing robots.txt sat in a grey zone with respect to the websites they were reading.
Reliability varied. Some plugins were operated by individual developers on small servers, and a backend outage simply broke the feature inside ChatGPT until the developer noticed. Users often kept two or three roughly equivalent internet plugins enabled at once.
The model itself was not always disciplined about when to call a plugin. Even with browsing enabled, ChatGPT would sometimes answer from memory and produce a stale answer. Good prompt engineering practice during the plugin era often included telling the model to use its internet plugin rather than relying on it to figure that out alone.
OpenAI announced the end of the plugin platform in early 2024. Two dates anchor the timeline:
| Date | Event |
|---|---|
| March 19, 2024 | New conversations with plugins were disabled. The plugin store stopped accepting new installs and new chats. |
| April 9, 2024 | Existing plugin conversations stopped working. The platform was effectively shut down. |
The replacement combined three things. Custom GPTs introduced on November 6, 2023, let builders package a system prompt, files, and external API calls (called "actions" rather than "plugins") into a shareable assistant. The GPT Store, opened January 10, 2024, gave custom GPTs a discovery layer similar to the old plugin store. OpenAI also rebuilt the most heavily used plugin functions, browsing in particular, as native features inside ChatGPT.
Many internet plugins survived in some form. WebPilot, AskYourPDF, ScholarAI, and others rebuilt as Custom GPTs, often porting the same backend API. Some companies, like Exa (formerly Metaphor), pivoted toward selling their search APIs directly to developers building their own large language model applications.
The internet category of ChatGPT plugins is interesting mostly for what it foreshadowed. It was an early, public, large-scale demonstration that a chat assistant could be combined with live web access through a simple tool interface, and that users would treat the result as a normal feature rather than a research demo. The same pattern reappears in almost every later product: function calling in the OpenAI API, tool use in other vendors' models, and agent frameworks that orchestrate web fetches and searches around an LLM.
The rapid sunset of the plugin platform also served as a case study in how quickly the surface area of an AI product can be reorganized. A category that did not formally exist before March 2023 was central to the ChatGPT user experience by mid-2023 and gone by April 2024, replaced by features built directly into the model's environment. Internet plugins are now a historical artifact of a thirteen-month window in which third-party developers, rather than OpenAI itself, were the primary suppliers of live web access for ChatGPT users.