To build a ChatGPT plugin in 2026, put a remote MCP server in front of your product, add OAuth 2.1 for user data, add optional MCP Apps UI, then upload a ZIP to OpenAI's plugin portal for review. A read-only plugin takes two to three engineer-weeks; one with UI and MCP Events, two to three months.
Updated 13 November 2026. "Plugin" is what OpenAI now calls a package of skills, an MCP server and optional UI that runs in ChatGPT and Codex. Since DevDay on 29 September 2026, plugins can also live in the ChatGPT sidebar, open panels beside a conversation, act as file viewers and start automations through MCP Events, and OpenAI's always-on dots reach your product through them. This guide covers the naming, the architecture, every build step, submission and review, cost and timeline, and what we found when we checked our own MCP servers against OpenAI's requirements. Every Apps SDK, directory and submission fact was re-checked on OpenAI's developer docs on 13 November 2026.
No. The original ChatGPT plugins from 2023 were a beta built on an OpenAPI manifest, and OpenAI shut them down in 2024. Today's ChatGPT plugins are a different product: a package of skills, a remote MCP server and optional UI, published to the ChatGPT directory and also usable in Codex.
The naming changed three times, which is why search results are confusing:
| Period | What OpenAI called it | How it worked | Status in November 2026 |
|---|---|---|---|
| 2023 to early 2024 | ChatGPT plugins (beta) | An ai-plugin.json manifest plus an OpenAPI spec of your REST API | Shut down. New plugin chats ended on 19 March 2024 and existing ones on 9 April 2024, per OpenAI's wind-down notice |
| Late 2023 onward | GPTs with Actions | A custom GPT that calls your API through an OpenAPI schema | A separate product, not covered here |
| 2025 to 2026 | Apps in ChatGPT, built with the Apps SDK | A remote MCP server with optional UI components | The Apps SDK documentation now redirects to the plugins docs |
| 2026 | Plugins for ChatGPT and Codex | Skills, MCP servers and optional UI in one package, with extensions and MCP Events | Current. Documented at developers.openai.com/plugins |
So if you search "how to create a plugin for ChatGPT", ignore tutorials that start with ai-plugin.json. Those describe the retired 2023 system. Everything in this guide follows the current plugin documentation.
Plugins went from tools the model calls in the background to app-like features inside ChatGPT. At DevDay on 29 September 2026, OpenAI announced:
OpenAI's DevDay 2026 recap puts ChatGPT at 1.2 billion weekly users, its own figure. A listing is not the same as being found, as the review section explains, but the plugin is now how your product shows up when a customer, or their dot, works inside ChatGPT. For how dots compare with building your own agent, see OpenAI Dots, Meta Muse or a custom AI agent.
A ChatGPT plugin is a package that contains skills, a remote MCP server, or both, with optional UI that the server returns. The MCP server is the core: ChatGPT, Codex and dots call its tools over Streamable HTTP, and it calls your API with the user's OAuth token. OpenAI's documentation describes four shapes:
| Shape | What it contains | Choose it when |
|---|---|---|
| Skills only | Folders with a SKILL.md and supporting files that tell the model how to run a workflow | Instructions plus tools ChatGPT already has are enough. No access to your product is needed. |
| MCP server only | A remote MCP server exposing your product's tools, with auth | The model needs live data or actions from your product. |
| Skills and MCP server | Workflow instructions that use your MCP tools | Users run multi-step workflows, such as "prepare the monthly report", on top of your tools. |
| MCP server with UI | The above plus UI resources rendered in ChatGPT | People need to inspect, compare, edit or confirm structured data visually. |
In words, the request path looks like this: the user (or a dot) asks ChatGPT something, the model picks one of your tools, ChatGPT calls your MCP endpoint with a bearer token, your server checks authorisation and calls your API, and the tool returns structured content. If the tool is linked to a UI resource, ChatGPT renders it in a sandboxed iframe that talks back over the MCP Apps bridge. For MCP Events, the direction reverses: your product POSTs signed webhooks to a callback URL ChatGPT gave you.
For a SaaS product, the MCP server defines which tools exist, their input and output schemas, who may call them, and what data they return. UI is optional, and OpenAI's guidance is to build the tools first and keep every tool useful without UI, so dots can run the same workflow headlessly. If you have not built an MCP server before, start with our MCP server development guide. It covers the protocol, OAuth 2.1 architecture and security in depth, so this article focuses on what is specific to ChatGPT.

Build a remote MCP server with a few well-annotated tools, add OAuth and optional UI, test it in ChatGPT, then package it as a ZIP and submit it in the plugin portal. Nine steps take you from idea to a published listing:
/mcp, on a stable public HTTPS domain. Keep a small set of tools with strict schemas and plain verb names like get_order_status. The current MCP revision is 2026-07-28; ChatGPT also still works with servers on the older initialize-based revisions.S256) and appends a resource parameter that you should copy into the token, usually as aud. It prefers Client ID Metadata Documents where your authorization server supports them, and also supports dynamic client registration or a predefined client. Declare auth per tool with securitySchemes, and return runtime errors carrying _meta["mcp/www_authenticate"] so ChatGPT shows the account-linking prompt.iss in every authorization response), ChatGPT uses a stable redirect URI. Otherwise it uses a callback-specific one. Copy the exact value from the plugin's management page into your allowlist.readOnlyHint, destructiveHint and openWorldHint to explicit true or false values. Review checks that they match real behaviour, and a justification in the form cannot override a wrong value on the server.text/html;profile=mcp-app, usually a ui:// URI) through _meta.ui.resourceUri, and the UI runs in an iframe that talks to ChatGPT over JSON-RPC on postMessage. Declare a narrow content security policy in _meta.ui.csp.plugin.json, skills, icons and MCP configuration, upload it in the portal, fix automated findings, then submit for review. The submission section below has the exact steps.For enterprise buyers, also request the openid and email scopes and expose a UserInfo endpoint. ChatGPT Enterprise can use the email domain to stop a corporate identity linking your plugin in a personal workspace.

MCP Events is a draft, experimental MCP extension that lets ChatGPT subscribe to things that happen in your product, such as a new message or a status change, and start a task when they occur. Your server receives subscriptions and delivers each matching event as a signed webhook. It is not the same as MCP notifications, which only flow over an open connection.
In OpenAI's MCP Events documentation, it works in Work chats on ChatGPT web, in the desktop app with Cloud selected, and with dots. The specification itself is still a design sketch in the MCP project's experimental repository, so expect changes. It is the most backend-heavy part of a plugin.
Your server advertises an events capability (in the server/discover response on the 2026-07-28 revision, or in the initialize response on older ones) and implements three methods on the same authenticated endpoint as your tools:
events/list describes each event, its filters and its payload schema.events/subscribe creates or refreshes a subscription. ChatGPT supplies a callback URL and a signing secret with a whsec_ prefix. Make creation idempotent.events/unsubscribe stops it.Then your server POSTs one signed event per request to the callback when something matches, with a unique eventId and an X-MCP-Subscription-Id header. The production details are what take time:
410 or 413.OpenAI's integration supports webhook delivery only, not polling or streaming. Its documentation also tells you to treat comments and other user-written text in events as data and never to put instructions for the model in the payload. That matters, because an event-triggered automation is an agent acting on text written by someone else. Our AI agent security checklist covers the controls around that.
Verify your identity and domain, upload the plugin ZIP in the OpenAI Platform's Plugins page, connect and scan the MCP server, add reviewer credentials and test cases, then submit for review. After approval you choose when to publish. These are the steps in OpenAI's submission guide as of 13 November 2026:
/.well-known/openai-apps-challenge on the MCP hostname or an eligible parent domain.If you are rejected, fix the feedback and submit a corrected ZIP, or reply to the rejection email to appeal.
Review checks who you are, whether the plugin works for a reviewer, and whether it follows OpenAI's data and commerce rules. Use this as a readiness checklist, based on the review requirements and plugin guidelines:
Three operational details catch teams out. Projects with EU data residency cannot submit plugins with MCP servers; use a global-residency project. The MCP server's origin (scheme, hostname or port) cannot change between versions, so plan your domain before the first submission; OpenAI's submission guide says to contact support if the URL must change. And deleted tools disappear quickly, while new or changed tools go live only after automated checks pass.
Finally, approval is not discovery. Approved plugins must be published manually, and enhanced distribution such as directory placement or proactive suggestions is reserved for plugins OpenAI judges to show strong real-world utility and user satisfaction. You cannot request it.

Yes, but not over the regulated data itself. OpenAI's guidelines bar public plugins from processing PHI and PCI-scoped card data. A healthtech or fintech product can still publish a plugin for scheduling, general product information, de-identified analytics or other non-regulated features. For workflows that touch regulated data, a private custom MCP connection inside a customer's workspace is the right route, with the model endpoint, contracts and audit trail scoped like the rest of your compliant system. Our guides to HIPAA-compliant AI agents and the fintech PCI DSS and SOC 2 checklist cover that side.
Plan two to three senior engineer-weeks for a read-only plugin over public data and eight to fourteen for an authenticated plugin with write tools and MCP Events, plus at least one review round. At market rates that is roughly $2.5K to $23K offshore or $8K to $168K at US mid-market rates.
Most of the cost is in authentication, events and review readiness, not the protocol. The table converts effort into money at two published market rate bands: the senior Asia rate of $31–$41 an hour from our 2026 Vietnam rates guide, and the $100–$300 an hour typical of US mid-market firms.
| Plugin scope | What's included | Effort (senior engineer-weeks) | At $31–$41/h | At $100–$300/h |
|---|---|---|---|---|
| Read-only, public data | MCP server with 5–10 tools, annotations, package, test cases, video, privacy review | 2–3 | $2.5K–$4.9K | $8K–$36K |
| Authenticated, read-only | Above plus OAuth 2.1 with PKCE, protected-resource metadata, client registration, per-tool auth, demo tenant | 4–7 | $5K–$11.5K | $16K–$84K |
| Authenticated with UI | Above plus MCP Apps UI, sidebar or panel extension, content security policy, UI testing | 6–10 | $7.4K–$16.4K | $24K–$120K |
| Write tools and MCP Events | Above plus write scopes, confirmations, idempotency, subscription store, signed webhooks, SSRF guard, abuse testing | 8–14 | $9.9K–$23K | $32K–$168K |
These are planning estimates, not quotes. They assume your API already exposes what the tools need and that the team knows OAuth. If your API does not exist yet, that work comes first, and our AI app development cost guide covers it. Review time is not in the table, because OpenAI says review timelines vary and it does not accept requests for expedited review. Plan for at least one round of feedback.
Three gaps, which are probably typical for servers built in 2025 or early 2026. BeevR runs three MCP servers: beevr.ai/mcp and ecocheck.ai/mcp (public, read-only) and ecocheck.ai/mcp/account (a customer's own data, OAuth 2.1 with PKCE). On 7 October 2026 we compared them with OpenAI's plugin documentation:
readOnlyHint, destructiveHint or openWorldHint. Every tool is read-only, so the fix was one line per server: all three servers now declare readOnlyHint: true, destructiveHint: false, idempotentHint: true and openWorldHint: false on every tool.events capability. OpenAI's current docs accept events on initialize-based servers too, so adding events does not force a protocol upgrade, but the subscription store, signing and SSRF work still apply.iss would mean proxying both the authorize and token steps. We chose the simpler route OpenAI documents: copy the exact redirect from the plugin's management page into the allowlist when we register the plugin.What already fits: protected-resource metadata, a 401 challenge that names the metadata and scope, S256 PKCE only, public clients, short-lived tokens with revocation, and read-only access enforced in the backend API rather than the MCP layer. If you are starting fresh, building to the current OpenAI docs from day one avoids all three gaps.
Not quite. The MCP server is the part that connects ChatGPT to your product. The plugin is the package you submit and users install, and it can also include skills, UI and metadata. A plugin can even be skills-only with no server.
No. UI is optional, and many good plugins return structured results only. Add UI where people need to compare, edit or confirm something visually, and keep every tool working without it so agents can use it in the background.
OpenAI does not publish a timeline and says requests for expedited review cannot be accommodated. Most rejections come from reviewers being unable to log in, failing test cases, undisclosed personal data or wrong annotations, so get those right before you submit.
Dots use the supported plugins a user has installed and enabled, so a published plugin is how your product becomes reachable to them. MCP Events also works with dots. Correct annotations matter: OpenAI says a dot's background research is read-only, and follow-up actions need the applicable permissions.
Yes. Both are MCP clients, and ChatGPT's UI is built on the open MCP Apps standard. Our EcoCheck account server already allows redirect URIs for both clients. Host-specific extras, such as ChatGPT sidebar extensions, need extra metadata but do not break other clients. OpenAI's docs also include a guide for converting a Claude Code plugin or connector into a plugin you can submit.
BeevR builds production MCP servers and the plugins on top of them, with OAuth, audit and review readiness designed in. We start from your three to five use cases, write the review test cases first, and ship a working demo every Friday. Pricing is fixed per phase, from $10K per phase, and you own the code from day one. See our AI agent development work.
Book a 30-minute call and tell us what you want ChatGPT users to do with your product.