OpenAI Apps SDK
Last reviewed
May 16, 2026
Sources
30 citations
Review status
Source-backed
Revision
v1 ยท 3,417 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 16, 2026
Sources
30 citations
Review status
Source-backed
Revision
v1 ยท 3,417 words
Add missing citations, update stale details, or suggest a clearer explanation.
| OpenAI Apps SDK | |
|---|---|
| Overview | |
| Type | Developer platform for in-chat applications |
| Developer | OpenAI |
| Announced | 6 October 2025 |
| Venue | OpenAI DevDay 2025, Fort Mason, San Francisco |
| Underlying protocol | Model Context Protocol (MCP) |
| Host surface | ChatGPT on Free, Go, Plus, and Pro plans |
| Initial availability | Logged-in ChatGPT users outside the European Economic Area, Switzerland, and the United Kingdom |
| Reach claimed at launch | More than 800 million weekly ChatGPT users |
| Companion releases | OpenAI AgentKit, GPT-5 Pro API, Sora 2 in API |
| Predecessor surfaces | ChatGPT Plugins (sunset April 2024), GPT Store (launched January 2024) |
| Monetization standard | Agentic Commerce Protocol, co-developed with Stripe |
| Status at launch | Preview SDK with pilot apps live; public submissions opened later |
OpenAI Apps SDK is a developer kit released by OpenAI at OpenAI DevDay on 6 October 2025 that lets third-party software run inside a ChatGPT conversation. The kit is built on top of the Model Context Protocol (MCP), the open standard for connecting language models to external data and tools, and extends it so a developer can ship not only backend logic but also the chat-embedded user interface, the authentication flow, and the conversational hooks that let ChatGPT decide when to invoke the app. The headline framing at launch was that ChatGPT, with its reported 800 million weekly users, was becoming a runtime for other people's software rather than a single chatbot that occasionally called out to plugins.
The Apps SDK shipped together with OpenAI AgentKit, GPT-5 Pro in the API, and Sora 2 in the API, but it was the announcement that drew the most coverage. Pilot apps from Booking.com, Canva, Coursera, Expedia, Figma, Spotify, and Zillow were live in the ChatGPT interface on day one for logged-in users outside the European Economic Area, Switzerland, and the United Kingdom. OpenAI announced a second wave of partners, including AllTrails, DoorDash, Instacart, OpenTable, Peloton, Target, theFork, and Uber, scheduled for late 2025. A public submission and review track and a directory were announced for later in the same year, with monetization via the Agentic Commerce Protocol, an open commerce standard co-developed with Stripe.
The Apps SDK is OpenAI's third attempt to turn ChatGPT into a developer surface, after the ChatGPT Plugins beta of 2023 (sunset in April 2024) and the GPT Store and Custom GPTs of late 2023 and early 2024. Unlike those predecessors, Apps SDK uses an open protocol that other companies can also implement, leans on a richer client-side UI model than plugins ever supported, and ships with explicit monetization plumbing.
The path from Plugins to Apps SDK runs through three distinct platform attempts in roughly thirty months, each addressing the shortcomings of the last.
ChatGPT Plugins were introduced in alpha in March 2023 as a way for ChatGPT to call third-party APIs described by an OpenAPI specification. Plugins were widely tried by developers but never crossed into mainstream use. The plugin store closed and new plugin chat creation was disabled on 19 March 2024, and existing plugin conversations stopped working on 9 April 2024. OpenAI's stated reason for the wind-down was that plugins had been superseded by Custom GPTs and Actions, which it considered a more flexible model.
Custom GPTs and the GPT Store launched in November 2023 and January 2024. A Custom GPT was a no-code persona configured with instructions, knowledge files, and optional Actions (OpenAPI-described tools, similar to plugin endpoints). The GPT Store opened the door for creators to publish and for users to discover GPTs by category, and OpenAI later added a creator revenue programme. Adoption was substantial in raw numbers (millions of GPTs were created) but the format was constrained: a Custom GPT could not run a rich client-side UI, the navigation paradigm was a separate GPT page rather than the main ChatGPT thread, and the experience often felt like a glorified system prompt with a few API calls bolted on.
By mid-2025, OpenAI faced two pressures. First, MCP, introduced by Anthropic in November 2024 and rapidly adopted across the industry, was becoming the de facto integration standard for AI agents. By autumn 2025 the protocol had been adopted by Google DeepMind, Microsoft, and thousands of developers, and the SDKs had crossed roughly 97 million downloads across Python and TypeScript. Second, the agent and shopping markets were drawing in heavyweight platforms, including Perplexity, Anthropic's Claude with web tools, and Google's Gemini with Workspace and Extensions. OpenAI's response was to converge on MCP itself and to use it as the substrate for a richer in-chat application format. The Apps SDK is that format.
An Apps SDK app is, at its core, an MCP server. The developer defines tools (functions the model can call), describes their inputs and outputs in JSON Schema, and hosts the server somewhere reachable by OpenAI. ChatGPT acts as the MCP client. When the model decides a tool is relevant during a conversation, it issues an MCP tool call, the server runs the request, and the response is returned to the model, which may then call further tools or produce a final assistant message.
Where Apps SDK extends standard MCP is on the user-facing side. A pure MCP server only returns text and structured data. An Apps SDK server can additionally return a widget, which is a piece of HTML and JavaScript that ChatGPT renders inline inside the conversation. The widget is sandboxed by the ChatGPT client and can host an interactive map, a player, a form, a card, a chart, or any other UI element the app needs. Widgets can receive events back from the model (for instance, when the user types a new request), and they can post events to the model (when the user clicks a button), which gives apps a two-way binding between conversation and embedded interface.
The Apps SDK build flow, as documented by OpenAI, has three phases. Plan: identify the use case, the tools, and the UI components the app needs. Build: stand up the MCP server, implement the tools, write the widgets, wire up authentication and any persistent state, and test in ChatGPT's developer mode. Deploy: host the server publicly, register the app with OpenAI, and connect it to ChatGPT. Authentication uses standard OAuth flows so the app can link a ChatGPT user to a Booking.com account, a Spotify profile, or a Zillow saved search.
Users reach an app in two ways. The first is direct mention by name. Typing "Spotify, make me a playlist for a weekend party" routes the request to the Spotify app. The second is automatic suggestion. ChatGPT examines the conversation, decides that a known app fits, and either calls the app silently or surfaces it as a recommendation. Automatic suggestion is what turns the directory from a list users have to remember into a live recommendation surface, and it is the feature that most clearly distinguishes Apps SDK from the older plugin model.
One practical artefact of the architecture is that an app can keep an interactive surface pinned in the conversation. The Coursera app, for example, pins a playing video to the top of the chat thread while the user continues typing below, so the lesson keeps running. The Zillow app renders an interactive map of listings that the user can pan, filter, and click without leaving ChatGPT. These behaviours are not possible inside a stateless plugin call.
Seven pilot apps were live on launch day, with a second wave announced for later in 2025. The table below lists what each app does inside ChatGPT, based on OpenAI's announcement and contemporaneous coverage.
| App | Category | What it does inside ChatGPT |
|---|---|---|
| Booking.com | Travel | Searches and books accommodation directly in chat with embedded property cards and dates |
| Canva | Design | Generates posters, slides, and social images from a text brief inside an interactive design widget |
| Coursera | Education | Plays lecture videos pinned to the top of the chat thread while the conversation continues |
| Expedia | Travel | Surfaces flights and hotels with itinerary previews and saved trip context |
| Figma | Design | Turns text outlines and rough sketches into editable diagrams or wireframes |
| Spotify | Music | Generates playlists, finds podcasts, and queues tracks against the user's Spotify account |
| Zillow | Real estate | Renders an interactive listings map with in-chat filters for price, beds, and area |
The second-wave partner list, communicated for late 2025, included AllTrails for trail discovery, DoorDash and Instacart for grocery and food delivery, OpenTable and theFork for restaurant booking, Peloton for fitness, Target for general retail, and Uber for rides and delivery. The mix suggests OpenAI was deliberately seeding categories that pair conversational planning with a transactional outcome: a trip, a meal, a workout, a purchase.
During the DevDay keynote demos, an OpenAI staffer asked ChatGPT through the Canva app to design a poster for a dog-walking business, and Canva returned several layouts inside the chat. Another demo asked Zillow for apartments within a budget near a neighbourhood and got back a live map. A third asked Spotify to put together a weekend party playlist, with the app opening directly in the dialogue thread rather than as a separate side panel.
At launch, only the seven pilot apps were live. OpenAI said it would begin accepting public submissions later in 2025, with an approval process and a directory. The directory was framed as a curated catalogue rather than a wide-open marketplace, in part to avoid the quality, spam, and trust problems that had affected earlier app surfaces such as the GPT Store.
Developers submit an app through a developer portal that asks for the production MCP server URL, the OAuth configuration if any, a metadata bundle (name, description, icon, categories, supported regions, sample prompts), and a sample widget rendering. OpenAI reviews the app for technical compliance with the MCP and Apps SDK specifications, for user-experience quality (does the widget render properly, does the app handle errors), and for policy compliance (no deceptive behaviour, no policy-violating content, appropriate handling of personal data). Approved apps appear in the in-product directory and can be invoked by name or surfaced via automatic suggestion.
OpenAI's documentation for Apps SDK is explicit that the platform is intended to be durable: the help center materials state there are no plans to deprecate the SDK, language that responds directly to developer concerns about how long earlier surfaces lasted.
Apps SDK shipped without an immediate revenue programme, but the monetization path was sketched on stage and in the announcement. OpenAI said app submissions and details on developer revenue would arrive later in 2025. The plumbing for in-chat purchases is the Agentic Commerce Protocol, an open standard co-developed with Stripe and announced jointly with the Apps SDK.
The Agentic Commerce Protocol defines how an AI agent (such as ChatGPT) can present a merchant's catalogue, take a user's selection, and pass payment information to a merchant in a way that is secure for the cardholder, traceable for the merchant, and standardised across AI clients. The protocol introduces Shared Payment Tokens, a Stripe mechanism by which a payment method captured by an AI surface can be securely forwarded to a merchant's existing payment processor. At launch Shared Payment Tokens supported Apple Pay, Google Pay, and standard card methods. A Stripe merchant could turn on agentic checkout in about a line of code.
In-product checkout inside ChatGPT, launched alongside Apps SDK, hit early adoption headwinds. Of Shopify's millions of merchants, only about a dozen enabled the native checkout in the first months. In February 2026, OpenAI announced it would scale back native checkout for general retail and instead route purchases through retailer apps such as Instacart, Target, and Expedia, while continuing to co-develop the Agentic Commerce Protocol with Stripe. The protocol itself remained the strategic bet, even as the in-chat checkout surface contracted.
For non-commerce apps, OpenAI was less specific about a revenue model at launch. The expectation, reading the announcement and follow-up commentary, was that paid apps and subscription tie-ins would be offered to developers when the directory opened to public submissions later in the year.
Apps SDK is one of several attempts in 2024 and 2025 to make AI chat surfaces a runtime for third-party software. The table below compares it to the most relevant predecessors and rivals.
| Platform | Vendor | Protocol | UI in chat | Submission model | Status late 2025 |
|---|---|---|---|---|---|
| OpenAI Apps SDK | OpenAI | MCP (open) | Yes, sandboxed widgets | Reviewed directory | Live for non-EEA logged-in users |
| ChatGPT Plugins | OpenAI | OpenAPI (proprietary) | No, text only | Plugin store (closed) | Deprecated April 2024 |
| GPT Store Custom GPTs | OpenAI | OpenAPI Actions | Limited | Open store | Active but lower priority |
| Gemini Extensions | Proprietary | Limited, mostly text | Curated, first-party only | Live, narrow set of services | |
| Claude with MCP tools | Anthropic | MCP (open) | Limited, primarily structured output | Tool integration, no app store | Live for paid and developer users |
| Perplexity native integrations | Perplexity | Proprietary | Yes, custom cards | Curated partner integrations | Live, smaller catalogue |
Gemini and Gemini Extensions sit in a different competitive position from Apps SDK. Gemini ships with deep, native integration to Google Workspace (Gmail, Docs, Sheets, Drive, Calendar, Maps) and to YouTube, all of which work without third-party plumbing. The trade-off is that the Extension catalogue is small and almost entirely Google first-party. Apps SDK, by contrast, leans on a third-party ecosystem and an open protocol; its breadth depends on how many outside developers it can attract.
The ChatGPT Plugins comparison is the most pointed one. Plugins were proprietary, text-only, and constrained to OpenAPI tools. Apps SDK is open-protocol, supports rich client-side UI, and ships with a commerce standard. Many of the structural reasons Plugins fizzled (poor discovery, weak monetization, inconsistent UX) are addressed in the Apps SDK design, although whether the new platform sustains user attention is a separate question.
The Claude comparison is interesting because Claude uses the same protocol. MCP servers built for Claude can in principle be reused for ChatGPT, but Apps SDK adds the widget extensions on top, so a fully featured ChatGPT app does more than a vanilla MCP integration. The protocol layer is shared; the app layer is not. The MCP Apps proposal published in early 2026 was a community effort to standardise the widget extensions across clients, which would erode the lock-in even further.
Reception of Apps SDK has been a mixture of strategic enthusiasm and tactical caution. The dominant narrative in October 2025 coverage from VentureBeat, TechCrunch, InfoQ, Latent Space, and others was that ChatGPT was attempting to become an app platform on the scale of iOS and the Web, with MCP as the open layer. VentureBeat went so far as to title its DevDay round-up "ChatGPT becomes the new app store". Founders and product leaders at the pilot apps said in OpenAI's announcement and on social media that conversion and engagement from the in-chat surface had exceeded their expectations during the closed pilot.
Developer reaction in the weeks after launch was more measured. Common observations on the OpenAI developer community forum and on independent blogs included confusion over the relationship between Apps SDK, AgentKit, Connector Registry, and standard MCP servers (which the documentation tried to clarify in later updates), a recurring 424 error code when widget responses were malformed, and uncertainty about the long-term commercial model for apps that were not directly transactional. Render, the application hosting provider, published a field guide in late 2025 walking through the practical pitfalls of building and hosting an Apps SDK server, including widget runtime quirks and webhook patterns.
The in-chat checkout walk-back in February 2026 was treated as a partial setback, not a strategy reversal. Coverage from Digital Commerce 360 and Rye argued that the underlying Agentic Commerce Protocol was the durable bet and that direct ChatGPT checkout had simply been the wrong first product to ship: retailer apps already had the catalogue, inventory, and fulfilment in place, and Apps SDK was a better fit for steering shoppers into those apps than for replacing them.
On the open-standard side, Apps SDK was widely read as a vote of confidence in MCP. The protocol's adoption by OpenAI, Anthropic, Google, and Microsoft, combined with the MCP donation to the Agentic AI Foundation in December 2025, established it as the de facto integration standard for the second half of the decade. Critics, summarised by Speakeasy in a sceptic-friendly explainer, raised the usual concerns about authentication gaps in MCP implementations, prompt injection risk through tool descriptions, and the maturity of the broader tooling. None of these criticisms were specific to Apps SDK, but they applied to it.
The longer-term question after launch was whether ChatGPT could sustain the kind of mainstream developer ecosystem that the iOS App Store had, or whether Apps SDK would settle into a smaller, transactional niche dominated by a few dozen large partners. The pilot list and the second wave were both heavy on consumer brands; whether a long tail of small developers would also commit remained open at the time of writing.