GPT-4 Plugins
Last reviewed
May 9, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v5 ยท 3,471 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 9, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v5 ยท 3,471 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: GPT-4, ChatGPT and Plugins
GPT-4 Plugins were a system of optional add-ons offered by OpenAI inside the ChatGPT product that let the underlying GPT-4 model call out to external services, retrieve live information, and perform real-world actions on behalf of a user. The feature was announced on March 23, 2023, opened in alpha to a waitlist of developers and a small group of subscribers, and rolled out more broadly to ChatGPT Plus customers through the spring of 2023. Plugins represented OpenAI's first attempt to turn ChatGPT into a platform that third parties could build on top of, and they directly shaped the design of later tool-use features such as the function calling API, custom GPTs, and the GPT Actions framework. The plugin system was officially deprecated on April 9, 2024, after roughly thirteen months of public availability, with custom GPTs and the GPT Store replacing it as the recommended way to extend ChatGPT.
A GPT-4 plugin was a thin connector between ChatGPT and an external API. Each plugin was described by a JSON manifest hosted on the developer's domain together with an OpenAPI specification that listed the available endpoints. When a user enabled a plugin in the ChatGPT interface, GPT-4 received a short natural-language description of what the plugin could do, the names and parameters of each endpoint, and the URL where requests should be sent. During a conversation, the model decided when to call a plugin, generated the appropriate request, parsed the response, and folded the result back into its reply. From the user's perspective the experience felt like a single chat with ChatGPT, but underneath the model was orchestrating one or more API calls per turn.
Plugins were available only to paying ChatGPT users. They appeared first inside ChatGPT Plus, then later in ChatGPT Enterprise after that tier launched in August 2023. Free-tier users never received access. The feature lived behind a settings toggle labelled "Beta features" and required users to opt in before they could browse and install plugins from the in-product plugin store.
GPT-4 was released by OpenAI on March 14, 2023 as the company's most capable large language model at the time. It was offered initially through ChatGPT Plus, which had launched in February 2023 at twenty United States dollars per month, and through the OpenAI API on a waitlist. GPT-4 was multimodal in that it could accept images as well as text input, although the image input feature for ChatGPT users did not ship until later in the year. The model showed marked improvements over GPT-3.5 on tasks that required multi-step reasoning, longer-form writing, and the kind of structured output that would later become important for tool use.
At launch, however, GPT-4 was still a closed system. Its responses were limited to its training data, which had a cutoff of September 2021, and it had no way to look anything up, run code, or take actions outside the chat window. OpenAI had been experimenting internally with giving language models access to tools for some time, drawing on academic work such as Toolformer and on its own earlier WebGPT prototype. The plugin announcement nine days after the GPT-4 launch was the first public version of that work.
On March 23, 2023, OpenAI published a blog post titled "ChatGPT plugins" introducing the system. The post described plugins as "eyes and ears" for ChatGPT and framed them as a way to address two long-standing limitations of the model: stale training data and the lack of any ability to take action in the world. OpenAI shipped two first-party plugins of its own, a web browsing plugin and a Codex-based code interpreter, alongside an open-source retrieval plugin that developers could self-host to expose private documents to ChatGPT. Eleven launch partners provided the first batch of third-party plugins: Expedia, FiscalNote, Instacart, KAYAK, Klarna, Milo, OpenTable, Shopify, Slack, Speak, Wolfram, and Zapier.
Access was gated behind a waitlist for both developers and end users. OpenAI explained the staged rollout as a way to study real-world use, gather safety feedback, and iterate on the developer interface before opening the gates more widely.
Every plugin was identified by a JSON file served at the well-known path /.well-known/ai-plugin.json on the developer's domain. The manifest contained the plugin's display name, a short description for end users, a longer description aimed at the model, the URL of an OpenAPI specification, an authentication scheme, contact details, and a logo. When ChatGPT installed a plugin, it fetched the manifest, then the OpenAPI document, then constructed a TypeScript-flavoured pseudocode summary of each endpoint that was injected into the system prompt for any conversation in which the plugin was active.
The OpenAPI specification was the contract between GPT-4 and the plugin's backend. It listed every callable endpoint, the HTTP method, query and body parameters, and the JSON shape of the response. Because OpenAPI is a public standard already used widely for documenting REST APIs, many companies could expose existing services to ChatGPT with relatively little new work.
A single user turn could trigger several plugin calls. When the user asked a question that touched a plugin's domain, GPT-4 would emit a structured request that ChatGPT's runtime would intercept, forward to the plugin's API, and then feed the response back to the model as a new message. The model would then decide whether the answer was complete, whether another plugin call was needed, or whether it should ask the user a clarifying question. This loop of model output, tool execution, and tool result is now usually described as tool use or agentic behaviour, but in early 2023 the plugin system was the first place where everyday ChatGPT users could see it in action.
Multiple plugins could be enabled simultaneously, with a soft cap of three active plugins per conversation. GPT-4 was responsible for routing questions to the right plugin or for combining results across plugins, for example by asking the Wolfram plugin for a numerical answer and then handing that number to the Zapier plugin to update a spreadsheet.
Plugins could declare one of several authentication modes in their manifest, ranging from no authentication for public APIs through static service-level tokens to OAuth 2.0 flows that linked a user's ChatGPT account to a third-party service. OpenAI ran a manual review process for every new plugin before it could appear in the in-product store. The browsing plugin was sandboxed to GET requests only to limit the blast radius of any prompt-injection attack, and the code interpreter plugin ran user-submitted Python in an isolated container with no internet access.
Despite these guardrails, the plugin system surfaced new safety problems that researchers and red-teamers documented during 2023. Indirect prompt injection through plugin responses, exfiltration of conversation context to third-party servers, and the risk of models being tricked into chaining benign-looking calls into harmful workflows all became active areas of study, and the lessons fed directly into the design of the later GPT Actions system.
| Date | Event |
|---|---|
| March 14, 2023 | GPT-4 released to ChatGPT Plus and to the OpenAI API waitlist |
| March 23, 2023 | OpenAI announces ChatGPT plugins; alpha access opens to a waitlist of developers and ChatGPT Plus subscribers |
| April 2023 | Web browsing and code interpreter ship to alpha testers; first third-party plugins go live for early users |
| May 12, 2023 | Plugin store reaches roughly 70 third-party plugins |
| May 13 to 14, 2023 | OpenAI begins rolling plugins out to all ChatGPT Plus subscribers, ending the user-side waitlist |
| June 1, 2023 | Plugin store contains around 210 plugins |
| June 13, 2023 | OpenAI ships the function calling API for GPT-4 and GPT-3.5, exposing the same tool-use capability to API developers |
| August 29, 2023 | Plugin store crosses roughly 940 plugins |
| August 28, 2023 | ChatGPT Enterprise launches, also gaining access to plugins |
| November 6, 2023 | OpenAI DevDay; OpenAI announces custom GPTs, the Assistants API, and signals that plugins will be wound down |
| January 10, 2024 | GPT Store launches publicly to ChatGPT Plus, Team, and Enterprise users |
| March 19, 2024 | Users can no longer start new conversations with plugins enabled |
| April 9, 2024 | Plugins beta is officially shut down; existing plugin conversations stop working |
OpenAI shipped three plugins under its own name during the lifetime of the system. These were the most-used plugins in practice and shaped user expectations about what the technology could do.
| Plugin | Purpose | Notes |
|---|---|---|
| Web browsing | Gave GPT-4 access to a text-only browser that could issue GET requests against the Bing search index, follow links, and quote sources | Inherited Microsoft's Bing safe-mode filtering; later evolved into the "Browse with Bing" mode and eventually into the integrated browsing in ChatGPT |
| Code interpreter | Provided a sandboxed Python environment with ephemeral disk space, allowing the model to run user-supplied or model-generated code, manipulate uploaded files, and produce charts | Renamed to Advanced Data Analysis in August 2023; survived the plugin sunset and remained part of ChatGPT under that name |
| Retrieval | Open-source plugin published on GitHub that developers could self-host to expose embeddings-based search over private documents to ChatGPT | Apache-2.0 licensed; widely used as a reference design for retrieval-augmented generation patterns |
OpenAI's launch partners, the small businesses that built plugins in the months that followed, and a few standout community-built plugins illustrate the breadth of the ecosystem.
| Plugin | Image | Platform | Category | Release Date | Description | Available | Working |
|---|---|---|---|---|---|---|---|
| BuyWisely (ChatGPT Plugin) |  | ChatGPT | E-Commerce | May 20, 2023 | Compare Prices & Discover the Latest Offers from thousands of online shops in Australia. | Yes | Yes |
| Instacart (ChatGPT Plugin) |  | ChatGPT | Food Delivery | May 20, 2023 | Ask about recipes and discover the ingredients you need, then get it all delivered from local stores. | Yes | Yes |
| Speechki (ChatGPT Plugin) |  | ChatGPT | Technology | May 20, 2023 | The easiest way to convert texts to ready-to-use audio - download link, audio player page, or embed! | Yes | Yes |
The original eleven launch partners covered a deliberately diverse set of categories so OpenAI could observe how GPT-4 handled different kinds of tools.
| Partner | Category | What the plugin did |
|---|---|---|
| Expedia | Travel | Searched flights, hotels, vacation rentals, activities, and car rentals using live availability and pricing |
| FiscalNote | Legal and policy data | Returned curated datasets covering legal, political, and regulatory information |
| Instacart | Grocery delivery | Ordered ingredients from local stores in the United States and Canada based on a chat-generated recipe |
| KAYAK | Travel | Returned flight, stay, and rental-car options matching constraints described in natural language |
| Klarna | Shopping and finance | Searched and compared prices on millions of products from online retailers |
| Milo | Family logistics | Provided a daily schedule and helper actions for parents of school-age children |
| OpenTable | Restaurant reservations | Searched availability and booked tables at participating restaurants |
| Shopify | E-commerce | Searched products listed on Shopify-powered stores worldwide |
| Slack | Team messaging | Summarised threads, drafted replies, and queried channels inside a connected Slack workspace |
| Speak | Language learning | Provided AI-mediated conversational practice in Spanish and other languages |
| Wolfram | Knowledge and computation | Exposed Wolfram Alpha's curated knowledge graph and the Wolfram Language for symbolic and numeric computation |
| Zapier | Automation | Triggered workflows across more than five thousand apps connected through Zapier's automation platform |
In addition to the launch cohort, several plugins became staples of the everyday ChatGPT-with-plugins workflow during 2023.
| Plugin | Category | What it did |
|---|---|---|
| Wolfram | Computation | Most-cited plugin for solving mathematics, physics, and unit-conversion problems that GPT-4 alone struggled with |
| WebPilot | Browsing | Free web fetcher that read full article text from a URL, often used as an alternative to OpenAI's own browsing plugin |
| Link Reader | Browsing | Read PDFs, images, and other documents from any URL and returned the text to GPT-4 |
| AskYourPDF | Documents | Allowed users to upload a PDF and ask questions about its contents |
| ScholarAI | Research | Searched a database of more than two hundred million academic papers and returned summaries with citations |
| Prompt Perfect | Prompting | Rewrote a user's prompt to be more specific before passing it to GPT-4 |
| Show Me Diagrams | Visualisation | Generated Mermaid and other diagrams from a natural-language description |
| VoxScript | Video and finance | Searched YouTube transcripts, retrieved stock data, and read web pages |
| ChatWithPDF | Documents | Loaded a PDF from a URL and answered questions about it |
| There's An AI For That | Discovery | Searched a directory of AI tools by use case |
OpenAI did not publish official statistics about the size of the plugin store, but third-party trackers and contemporary press reports captured the rough trajectory.
| Date | Approximate plugin count |
|---|---|
| March 23, 2023 | 14 (eleven launch partners plus three OpenAI plugins) |
| May 12, 2023 | 70 |
| June 1, 2023 | 210 |
| August 29, 2023 | 940 |
| October 2023 | More than 1,000 |
By late 2023 the store had crossed roughly one thousand plugins, but usage data shared by OpenAI at DevDay and in the deprecation post indicated that activity was concentrated in a small number of plugins. Most ChatGPT Plus subscribers never enabled any plugin, and even among power users the typical session relied on a handful of favourites such as Wolfram, browsing, and code interpreter.
On June 13, 2023 OpenAI shipped a feature in the API that exposed the same underlying tool-use capability to developers without requiring the full plugin manifest pipeline. The new gpt-4-0613 and gpt-3.5-turbo-0613 models accepted a list of function descriptions in each request and could choose to respond with a structured JSON object indicating which function to call and with what arguments. Function calling quickly became the canonical way to give a large language model access to external tools through the API, and it survived multiple model generations as a stable interface.
At OpenAI DevDay on November 6, 2023, Sam Altman introduced custom GPTs, no-code agents that any ChatGPT subscriber could build by writing a system prompt, uploading reference files, and optionally connecting an external API through a feature called Actions. Actions inherited the OpenAPI-based design of plugins but lived inside an individual GPT rather than in a shared store, and they could be combined with the same browsing, code interpreter, and image generation tools that ChatGPT exposed natively.
The public GPT Store launched on January 10, 2024, giving creators a curated catalogue, search and discovery, and eventually a revenue-sharing program based on usage. From the user's perspective a custom GPT was a single configurable agent rather than a toggleable plugin, which OpenAI argued was a less confusing model than asking users to assemble a toolkit themselves.
The Assistants API, also announced at DevDay 2023, brought stateful conversations, file handling, retrieval, code interpreter, and function calling together inside a single developer-facing surface. It removed much of the boilerplate that early plugin developers had written to manage conversation state and tool orchestration, and it became the recommended way to build agentic applications on top of GPT-4 and its successors.
On February 23, 2024 OpenAI published a help-centre article announcing that the plugins beta would be wound down. The post explained that GPTs offered a better path to reach ChatGPT users, both because they were easier for end users to discover and because they let creators package an entire experience rather than a single tool. The deprecation followed a two-stage schedule. From March 19, 2024 onward, no new conversations could be started with plugins enabled, although existing plugin chats remained accessible. On April 9, 2024 the second stage closed those existing conversations and removed the plugin store from the ChatGPT interface. Developers were advised to port their plugins to GPT Actions, which retained the same OpenAPI-based contract.
Public statements from OpenAI and analysis at the time pointed to several reasons for the shutdown. Plugin discovery was poor: the in-product store had no recommendations, no usage signals, and no editorial curation, so most subscribers never found anything worth installing. The three-plugin cap and the per-conversation activation flow meant that combining tools required planning by the user, which most casual users were unwilling to do. The plugin format also lagged the rapid evolution of the underlying models; features such as long context windows, image input, and persistent memory could not be exposed through the static plugin contract without major redesign. Custom GPTs absorbed all of those concerns by bundling tools, instructions, and knowledge into a single shareable artifact.
The plugin system left a durable mark on how language-model products are built. The OpenAPI-based contract, the manifest-and-spec packaging model, and the runtime loop in which the model decides when to call a tool all flowed forward into function calling, Assistants, GPT Actions, and competing systems from other vendors. Plugins also gave the broader public its first exposure to the idea of an LLM that could browse, calculate, and act, an idea that became central to the agent boom of 2024 and 2025.
Reaction to the plugin launch in March 2023 was unusually positive even by the standards of OpenAI announcements. Press coverage at TechCrunch, VentureBeat, and other outlets framed plugins as ChatGPT's "App Store moment" and emphasised that the system shipped with strong launch partners covering travel, shopping, and productivity. Industry analysts pointed to the use of OpenAPI as a savvy choice that lowered the barrier for established companies to participate.
User sentiment cooled somewhat over the course of 2023. Forum threads on the OpenAI Developer Community and Reddit complained about slow plugin response times, frequent timeouts, the awkwardness of activating plugins through the settings menu, and the confusion of choosing among many similar plugins. Developers expressed frustration that their plugins received little organic discovery and that OpenAI's curation was inconsistent. By the time custom GPTs were announced at DevDay, much of the developer community had already concluded that plugins were a transitional design.