ChatGPT apps and plugins connect outside tools and services to ChatGPT, so people can use them during a conversation. For a software business, this can give customers another way to use its product. Some eligible integrations also let customers use their existing ChatGPT plan for AI requests, but that access has conditions. Plugin overview, Plan access
This guide explains who can use these connections, what they provide and which costs remain. Here, “OpenAI marketplace” refers to the ChatGPT app and plugin ecosystem, not a separate model API.
Who is this for?
This guide is for founders considering a ChatGPT connection for their product and developers deciding how to build it. It also helps customers understand whether an existing subscription can cover the AI features of another app.
A few terms will make the rest easier to follow:
- SaaS: software offered as an online service, often through a subscription.
- API: a way for one piece of software to request data or actions from another.
- Integration: the connection that lets two services work together.
- Authentication: checking who a user is; authorization determines what they may access.
- Inference: the work an AI model does to produce a response.
What can the connection provide?
| Capability | What it could help a customer do | What must be checked |
|---|---|---|
| Product tools inside a conversation | Use supported actions from your service | What your integration exposes and is approved to do |
| ChatGPT sign-in | Identify or connect their account | Access to the relevant sign-in route and requested permissions |
| Eligible ChatGPT-plan usage | Run supported AI requests using their plan | User eligibility, integration access and usage conditions |
These are separate capabilities. Publishing an integration should not be treated as automatic approval for subscription access or a payment flow.
Start with the job your integration will do
OpenAI's current developer documentation describes plugins combining skills, MCP servers and optional interfaces to connect services to ChatGPT and Codex. It includes development, testing and publishing guidance. MCP is a protocol for exposing tools and data; it is not itself a payment arrangement. Plugin documentation
For a founder, the first design question is practical: which task becomes easier inside a conversation?
Consider a hypothetical scheduling product. A useful integration could find an available appointment and help a customer book it. A less useful integration might simply repeat the homepage and send the customer elsewhere. The former has a concrete reason to exist, although its availability and approval would still require checking.
Write that task in one sentence before choosing an SDK: “A customer can check availability and request an appointment without manually searching our calendar.” That sentence gives you something to evaluate beyond whether a demo looks impressive.
The three connections are different
| Connection | Question it answers |
|---|---|
| Your service exposed as tools | What can ChatGPT do with your product? |
| Sign in with ChatGPT | Who is the user, and what have they authorized? |
| Eligible model requests | Which route runs the AI request, and whose allowance covers it? |
An integration can involve more than one of these. Do not assume that implementing one automatically enables the others.
Can a customer bring their ChatGPT plan?
| User or developer | Availability described in the official documentation |
|---|---|
| Eligible Plus and Pro users | Plan usage in participating apps |
| Commercial partners | Sign-in through a limited trial for selected partners |
| Open-source partners and selected private clients | Routes for ChatGPT-plan usage |
| Identity-only sign-in | Identifies the user; does not itself grant model access |
Model access requires separate authorization for eligible Responses API requests. The connection does not share the user's API key or conversations. Sign in with ChatGPT quickstart
That is a meaningful capability, but it is not a universal entitlement to free API usage for every commercial SaaS. Check the applicable integration route and its limitations before promising customers that their subscription will cover requests.
How would it work in practice?
For a hypothetical scheduling service, a customer asks for an available appointment, connects an account where needed, and uses the supported booking action. The precise permissions and confirmation steps depend on the integration. This example illustrates the workflow; it is not a tested or approved product.
For planning, draw the request path:
Customer action → your integration → your backend or eligible model endpoint → result.
At each step, identify authentication, usage limits, storage and cost.
What costs and limits should you check?
Even where eligible inference uses a customer's plan, your own hosting, database, support and third-party services still need a budget. Check these items before quoting a price:
- Which requests qualify for the plan-backed route, and what happens when access is unavailable.
- Any separate model requests your backend makes and how those are billed.
- Limits or charges from your own product and other connected services.
- The applicable account, regional and preview restrictions.
This guide does not establish one allowance or price that applies to every integration. Consult the documentation for the route you can actually use.
Marketplace presence does not settle monetization
OpenAI's current checkout reference recommends merchant-hosted external checkout for eligible purchases. It says current approval is limited to plugins for physical-goods purchases, with embedded payment-sheet access limited to selected marketplace partners. These provisions should not be read as approval to sell any SaaS subscription through an embedded checkout. Checkout documentation
Before designing a payment flow, confirm what your product category and integration are allowed to do. Distinguish a purchase made through your existing service from a new purchase initiated inside the platform.
Build your forecast from a supported route. A payment feature on a roadmap is not revenue you can rely on today.
Assess the distribution opportunity realistically
Our editorial view is that a conversational integration deserves attention when it removes a real step from a customer's work. Its presence alone does not prove that customers will discover it, activate it or keep using it.
For an initial experiment, measure a small sequence: people who find the integration, connect an account, complete the intended task and return. Where you can measure those stages appropriately, they will help separate curiosity from useful adoption.
Also keep your existing distribution working. Documentation, onboarding, search visibility and customer relationships remain assets you control more directly than another platform's discovery interface.
Practical steps before you build
First choose one customer task, then check its fit against the official integration documentation. Confirm your access route before designing around subscription usage. Map permissions, requests and costs, build a small test, and follow the applicable testing and submission process. OpenAI documents those stages in its plugin development guide.
Use the following questions to review your plan:
- Which customer task belongs in a conversation?
- Can it be completed safely with the permissions the customer grants?
- Is our integration route currently available to our type of business?
- Who pays for each model request and each external service?
- What happens when the allowance runs out or the service fails?
- Is the intended payment flow explicitly supported?
If the answers are clear, you have the basis for a bounded experiment. If they are still assumptions, resolve them before promising a launch date or subscription benefit.
Research note: This is a documentation-based explanation, not a tested integration or a guarantee of listing, approval or revenue. Product terminology and access conditions can change; verify them before implementation.