How to Build a ChatGPT Plugin Without an Existing API

You need a usable source, not necessarily a new business platform

You can build a ChatGPT plugin without an API in your own business software if you have another authorised way to obtain the required information. A managed spreadsheet, an approved export or a supported adapter may be enough for a bounded task. Live actions require a connection that can reliably perform and verify them. First distinguish two situations. Your business may not have built an API, but a product it already uses may offer one. Alternatively, the source may genuinely provide only files or a staff interface. Those situations deserve different estimates. This article proposes a connection ladder for a business that wants customers to check service scope. It does not assume that all legacy systems can safely become real-time booking systems.

Decide how current the answer must be

Write the freshness requirement before choosing a connector. A service description may be acceptable until the business changes its offering. Appointment capacity may change whenever staff accept another job. For each proposed capability, state what happens when the source is too old or cannot be read. A last-approved catalogue can support a general explanation, but it should not imply that a particular slot is still available. Consider an example with a nightly export at 7 pm. At 10 am the next day, the snapshot is 15 hours old. That might be acceptable for a stable catalogue and unsuitable for current capacity. The age is arithmetic; acceptability is a business decision.

First rung: approved reference material

If the task explains stable scope or preparation requirements, approved reference material may be sufficient. Give the information an owner, a review process and explicit limits. A reusable instruction workflow may not need a live server at all. OpenAI's plugin architecture allows skills as well as server-backed capabilities. Choose the component that the task needs. Reference material cannot verify a changing customer record merely because it contains a sample record. Use this rung for questions whose answer can be maintained deliberately. If customers need private history or current availability, move to an appropriate live connection rather than adding a stronger instruction to an old document.

Second rung: a maintained spreadsheet

A spreadsheet can work for a small catalogue if the team controls its structure and access. Use stable identifiers and defined fields. Avoid relying on row positions that change when someone sorts the sheet. Assign the editor and approval process. Separate public catalogue fields from private notes. Decide whether an unapproved edit should become visible immediately or after a deliberate release step. If the implementation uses Google Sheets, account for its API quotas and backoff guidance. The current documentation lists read quotas of 300 requests per minute per project and 60 per minute per user per project. Design reads around the actual workload and authorised access rather than assuming unlimited capacity.

Third rung: a scheduled export

An approved CSV or file export can support a read-only snapshot. Record the export time, validate required fields and replace the active snapshot only when the new file passes checks. For the service catalogue, check that identifiers are unique, categories are recognised and coverage values are valid. An empty export should be investigated rather than automatically interpreted as a business offering no services. Keep the last accepted version available for recovery, with the agreed freshness rules. Explain when the data was current. The customer should not have to infer whether the result came from live software or yesterday's file.

Fourth rung: a supported adapter

A product may offer an API, database view or integration route that your team has not used. Ask the vendor what it supports and whether your plan includes it. Confirm the permitted use rather than treating access to a staff screen as permission to extract everything behind it. A custom server can translate a supported source into focused MCP tools. That can preserve the existing system while providing a simpler customer capability. The server build guide explains the interface the plugin exposes. Estimate the adapter separately from the customer-facing tool. A source with poor identifiers or inconsistent statuses can require more mapping work than a clean API, even when both return records successfully.

Treat browser automation as a deliberate exception

If the only route is a staff interface, assess whether authorised browser automation is appropriate. Check the vendor's access rules, login requirements and operational risks. A screen that can be operated manually is not automatically a reliable public integration. Start with a bounded task and a way to detect failure. A changed label, expired session or unexpected modal can alter behaviour. A read-only trial is different from submitting or cancelling customer work. If the proposed operation cannot be verified after an interruption, reconsider it. For a booking change, the business needs to know whether the change happened before retrying. A human fallback may be more appropriate than pretending the interface behaves like a transactional API.

Keep information retrieval separate from actions

A read-only catalogue can be useful without taking bookings. Do not add writes merely because the demo can fill a form. An action needs an accepted destination, outcome verification and duplicate handling. The CRM connection article explains why a successful write can still produce an uncertain customer result. Apply that reasoning to any source route, including automation through an interface. For a first pilot, define an information outcome the business can verify. If the answer reveals useful demand for an action, scope the action with the relevant source and eligibility requirements afterwards.

Evaluate a no-code product against the real task

Ask what the product connects to, how information is updated and how account access works. Check whether it supports the current ChatGPT plugin route you intend to use. An older tutorial using a discontinued manifest is not evidence of today's compatibility. Have the vendor demonstrate a changed source record, an unavailable source and a user without permission. Those cases reveal more than a well-prepared question against a static file. Review ownership and exit options. Can the business export its source information and operating instructions? What happens if the provider changes the connector? A lower initial build effort can still introduce ongoing dependence that needs pricing.

Questions about missing APIs

Can we start with CSV files? For a suitable read-only task, provided the export is authorised, validated and current enough for the answer it supports. Does using a no-code tool remove integration work? It can reduce implementation effort. Source ownership, access, freshness and recovery still need decisions. Should we rebuild our whole system first? Assess the smallest supported connection that fulfils the customer task. A full rebuild is a separate business decision.

Choose the lowest rung that does the job

Documentation checked on 5 October 2026. The ladder and catalogue example are proposed designs. Discuss your source system with 13labs before deciding whether you need a new API, an adapter or a more limited first plugin.

How to Build a ChatGPT Plugin Without an Existing API

Contents

You need a usable source, not necessarily a new business platform

Decide how current the answer must be

First rung: approved reference material

Second rung: a maintained spreadsheet

Third rung: a scheduled export

Fourth rung: a supported adapter

Treat browser automation as a deliberate exception

Keep information retrieval separate from actions

Evaluate a no-code product against the real task

Questions about missing APIs

Choose the lowest rung that does the job

Build a useful plugin for your business

Related Reading

How to Build a ChatGPT Plugin for Your Business

How Much Does a Business ChatGPT Plugin Cost to Build?

How to Let Customers Check Order or Booking Status in ChatGPT